remove duplicate save buttons

Co-authored-by: Cursor <cursoragent@cursor.com>

.claude/skills/crud-endpoints/SKILL.md

@@ -0,0 +1,31 @@
+---
+name: crud-endpoints
+description: Use whenever asked to add, create, or scaffold a CRUD endpoint, router, or entity in this repo's server (create/list/get/update/delete handlers, new `server/routers/<entity>/` or `server/private/routers/<entity>/` folder). Points to the established file layout, middleware, ActionsEnum, and route-registration conventions before writing any code.
+---
+
+Before writing any router/handler/middleware code for a new entity, read
+`docs/crud-endpoints.md` in full. It documents, with real examples from
+`server/routers/aiProvider/` (public) and `server/private/routers/alertRule/`
+(enterprise-only), how this repo structures CRUD endpoints:
+
+- Directory/file layout per entity (`index.ts`, `types.ts`, `validation.ts`,
+  one file per operation).
+- The standard handler anatomy (zod parsing, OpenAPI registry, response
+  envelope, error handling).
+- Where access-control middleware (`verify<Entity>Access`) lives and when
+  it's needed vs. plain `verifyOrgAccess`.
+- How to wire up `ActionsEnum` entries, `verifyUserHasAction`, and
+  `logActionAudit`.
+- Which of the four router files (`server/routers/external.ts`,
+  `server/routers/internal.ts`, `server/private/routers/external.ts`,
+  `server/private/routers/internal.ts`) to register routes in, and the
+  middleware chain template per HTTP verb.
+- The repo's non-standard verb convention: **`PUT` = create, `POST` =
+  update** (backwards from typical REST) — don't "fix" this to standard
+  REST verbs, match the existing convention.
+- The `#dynamic` import alias, for the rare case of a hook needing different
+  implementations in OSS vs. enterprise builds.
+
+Follow that doc's checklist (§8) step by step rather than improvising a
+structure. If the doc and the actual code in `aiProvider`/`alertRule` ever
+disagree, trust the code and flag the doc as stale.

Dockerfile

@@ -1,5 +1,5 @@
-# FROM node:24-slim AS base
-FROM public.ecr.aws/docker/library/node:24-slim AS base
+# FROM node:24.18.1-slim AS base
+FROM public.ecr.aws/docker/library/node:24.18.1-slim AS base
 
 WORKDIR /app
 
@@ -32,8 +32,8 @@ FROM base AS builder
 
 RUN npm ci --omit=dev
 
-# FROM node:24-slim AS runner
-FROM public.ecr.aws/docker/library/node:24-slim AS runner
+# FROM node:24.18.1-slim AS runner
+FROM public.ecr.aws/docker/library/node:24.18.1-slim AS runner
 
 WORKDIR /app
 

Dockerfile.dev

@@ -1,4 +1,4 @@
-FROM node:24-alpine
+FROM node:24.18.1-alpine
 
 WORKDIR /app
 

cli/commands/rotateServerSecret.ts

@@ -1,5 +1,5 @@
 import { CommandModule } from "yargs";
-import { db, idpOidcConfig, licenseKey, certificates, eventStreamingDestinations, alertWebhookActions } from "@server/db";
+import { db, idpOidcConfig, licenseKey, certificates, eventStreamingDestinations, alertWebhookActions, aiProviders } from "@server/db";
 import { encrypt, decrypt } from "@server/lib/crypto";
 import { configFilePath1, configFilePath2 } from "@server/lib/consts";
 import { eq } from "drizzle-orm";
@@ -132,12 +132,14 @@ export const rotateServerSecret: CommandModule<
             const certs = await db.select().from(certificates);
             const streamingDestinations = await db.select().from(eventStreamingDestinations);
             const webhookActions = await db.select().from(alertWebhookActions);
+            const providers = await db.select().from(aiProviders);
 
             console.log(`Found ${idpConfigs.length} OIDC IdP configuration(s)`);
             console.log(`Found ${licenseKeys.length} license key(s)`);
             console.log(`Found ${certs.length} certificate(s)`);
             console.log(`Found ${streamingDestinations.length} event streaming destination(s)`);
             console.log(`Found ${webhookActions.length} alert webhook action(s)`);
+            console.log(`Found ${providers.length} AI provider(s)`);
 
             // Prepare all decrypted and re-encrypted values
             console.log("\nDecrypting and re-encrypting values...");
@@ -171,11 +173,18 @@ export const rotateServerSecret: CommandModule<
                 encryptedConfig: string;
             };
 
+            type AiProviderUpdate = {
+                providerId: number;
+                encryptedApiKey: string | null;
+                encryptedHeaders: string | null;
+            };
+
             const idpUpdates: IdpUpdate[] = [];
             const licenseKeyUpdates: LicenseKeyUpdate[] = [];
             const certUpdates: CertUpdate[] = [];
             const streamingDestinationUpdates: StreamingDestinationUpdate[] = [];
             const webhookActionUpdates: WebhookActionUpdate[] = [];
+            const aiProviderUpdates: AiProviderUpdate[] = [];
 
             // Process idpOidcConfig entries
             for (const idpConfig of idpConfigs) {
@@ -306,6 +315,37 @@ export const rotateServerSecret: CommandModule<
                 }
             }
 
+            // Process aiProviders entries (apiKey + headers)
+            for (const provider of providers) {
+                try {
+                    if (!provider.apiKey && !provider.headers) {
+                        continue;
+                    }
+
+                    const encryptedApiKey = provider.apiKey
+                        ? encrypt(decrypt(provider.apiKey, oldSecret), newSecret)
+                        : null;
+                    const encryptedHeaders = provider.headers
+                        ? encrypt(
+                              decrypt(provider.headers, oldSecret),
+                              newSecret
+                          )
+                        : null;
+
+                    aiProviderUpdates.push({
+                        providerId: provider.providerId,
+                        encryptedApiKey,
+                        encryptedHeaders
+                    });
+                } catch (error) {
+                    console.error(
+                        `Error processing AI provider ${provider.providerId}:`,
+                        error
+                    );
+                    throw error;
+                }
+            }
+
             // Perform all database updates in a single transaction
             console.log("\nUpdating database in transaction...");
             await db.transaction(async (trx) => {
@@ -376,6 +416,17 @@ export const rotateServerSecret: CommandModule<
                             )
                         );
                 }
+
+                // Update AI provider entries
+                for (const update of aiProviderUpdates) {
+                    await trx
+                        .update(aiProviders)
+                        .set({
+                            apiKey: update.encryptedApiKey,
+                            headers: update.encryptedHeaders
+                        })
+                        .where(eq(aiProviders.providerId, update.providerId));
+                }
             });
 
             console.log(`Rotated ${idpUpdates.length} OIDC IdP configuration(s)`);
@@ -383,6 +434,7 @@ export const rotateServerSecret: CommandModule<
             console.log(`Rotated ${certUpdates.length} certificate(s)`);
             console.log(`Rotated ${streamingDestinationUpdates.length} event streaming destination(s)`);
             console.log(`Rotated ${webhookActionUpdates.length} alert webhook action(s)`);
+            console.log(`Rotated ${aiProviderUpdates.length} AI provider(s)`);
 
             // Update config file with new secret
             console.log("\nUpdating config file...");
@@ -402,6 +454,7 @@ export const rotateServerSecret: CommandModule<
             console.log(`  - Certificates: ${certUpdates.length}`);
             console.log(`  - Event streaming destinations: ${streamingDestinationUpdates.length}`);
             console.log(`  - Alert webhook actions: ${webhookActionUpdates.length}`);
+            console.log(`  - AI providers: ${aiProviderUpdates.length}`);
             console.log(
                 `\n  IMPORTANT: Restart the server for the new secret to take effect.`
             );

docs/crud-endpoints.md

@@ -0,0 +1,347 @@
+# How to build a CRUD endpoint in this repo
+
+Reference for adding a new CRUD entity to the server. Based on two real
+examples already in the codebase — read them side by side with this doc:
+
+- **Public / open-source (Community Edition) pattern**: `server/routers/aiProvider/`
+- **Enterprise-only pattern**: `server/private/routers/alertRule/`
+
+The two are structurally identical. The only difference is *where the files
+live* and *which router they get wired into*.
+
+## 1. Decide: public or private?
+
+- `server/routers/<entity>/` — ships in the open-source Community Edition.
+  Anyone running Pangolin gets this.
+- `server/private/routers/<entity>/` — Enterprise/SaaS only. Gated behind
+  `verifyValidLicense` (and often `verifyValidSubscription(tierMatrix.x)`).
+  Every file here starts with the Fossorial Commercial License header block
+  (copy it verbatim from an existing private file).
+
+Everything below applies to both — swap `@server/...` for `#private/...`
+import paths and add license headers when building the private version.
+
+## 2. Directory layout
+
+One folder per entity, one file per operation, a barrel `index.ts`:
+
+```
+server/routers/<entity>/
+  index.ts          # export * from each operation file + ./types
+  types.ts          # response payload types + row->public mapper
+  validation.ts      # zod schemas/refinements shared by create + update (optional)
+  create<Entity>.ts
+  list<Entities>.ts
+  get<Entity>.ts
+  update<Entity>.ts
+  delete<Entity>.ts
+```
+
+`index.ts` is a flat barrel:
+
+```ts
+export * from "./createAiProvider";
+export * from "./listAiProviders";
+export * from "./getAiProvider";
+export * from "./updateAiProvider";
+export * from "./deleteAiProvider";
+export * from "./types";
+```
+
+## 3. Anatomy of a single handler
+
+Every handler file (`create<Entity>.ts`, etc.) follows the same shape:
+
+```ts
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { <table>, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { eq } from "drizzle-orm";
+import type { GetXResponse } from "@server/routers/<entity>/types";
+
+const paramsSchema = z.strictObject({
+    orgId: z.string().nonempty()          // or entityId: z.coerce.number().int().positive()
+});
+
+const bodySchema = z.strictObject({ /* ... */ }); // create/update only
+
+registry.registerPath({
+    method: "get", // put | post | delete
+    path: "/org/{orgId}/x",
+    description: "...",
+    tags: [OpenAPITags.<Entity>],
+    request: { params: paramsSchema, /* body: {...} for write ops, query: for list */ },
+    responses: { 200: { description: "Successful response" } }
+});
+
+export async function getX(req: Request, res: Response, next: NextFunction): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(createHttpError(HttpCode.BAD_REQUEST, fromError(parsedParams.error).toString()));
+        }
+        // parse body too, if present, same pattern
+
+        // ...business logic against db...
+
+        if (!row) {
+            return next(createHttpError(HttpCode.NOT_FOUND, `X with ID ${id} not found`));
+        }
+
+        return response<GetXResponse>(res, {
+            data: { /* ... */ },
+            success: true,
+            error: false,
+            message: "X retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred"));
+    }
+}
+```
+
+Rules to keep consistent with the rest of the codebase:
+
+- `z.strictObject` for params/body — rejects unknown keys.
+- Params parsed first, then body; each on its own `safeParse` + early
+  `next(createHttpError(...))` — never throw raw errors.
+- Every handler registers itself with the OpenAPI `registry` even if nobody
+  reads the spec directly — it's how `/api/v1/docs` stays accurate.
+- Catch-all `try/catch` at the bottom: `logger.error(error)` +
+  generic `500` message. Never leak internal error details to the client.
+- Use `response<T>(res, { data, success, error, message, status })` from
+  `@server/lib/response` for every response, success or otherwise (errors go
+  through `next(createHttpError(...))` instead, not through `response`).
+- If the route already ran an access-control middleware that fetched the row
+  (see §5), reuse it instead of re-querying:
+  `req.aiProvider && req.aiProvider.providerId === providerId ? [req.aiProvider] : await db.select()...`
+
+### List handler specifics
+
+Pagination is a fixed shape (`page`, `pageSize`, optional `query` for
+search). See `listAiProviders.ts`:
+
+```ts
+const listSchema = z.object({
+    pageSize: z.coerce.number<string>().int().positive().optional().catch(20).default(20),
+    page: z.coerce.number<string>().int().min(0).optional().catch(1).default(1),
+    query: z.string().optional()
+});
+```
+
+Run the count query and the page query in `Promise.all`, and return
+`PaginatedResponse<{ items: T[] }>` (`@server/types/Pagination`) with
+`{ total, pageSize, page }`.
+
+### types.ts specifics
+
+- Define one response type per operation: `List<Entities>Response`,
+  `Get<Entity>Response`, `CreateOrEdit<Entity>Response` (create and update
+  commonly share a response shape).
+- If the raw DB row needs to be shaped for clients (decrypting secrets,
+  parsing a serialized column, hiding a column), put a `toPublic<Entity>()`
+  mapper here — see `toPublicAiProvider` for the pattern of stripping
+  `apiKey`/serialized columns and re-adding decrypted/parsed versions.
+
+### validation.ts specifics
+
+Only needed when create and update share non-trivial zod pieces (enums,
+`superRefine` cross-field rules). Export the raw schemas (`z.enum([...])`)
+and refinement functions, and import them into both `createX.ts` and
+`updateX.ts` — see `aiProvider/validation.ts`'s
+`refineProviderUpstreamFields`.
+
+## 4. Wire up an access-control middleware (for id-scoped routes)
+
+For routes scoped to a single row (`/x/:xId`, as opposed to
+`/org/:orgId/x` create/list), add a `verify<Entity>Access` middleware in
+`server/middlewares/` (or `server/private/middlewares/` for enterprise-only
+entities) and export it from that directory's `index.ts`.
+
+Pattern (`verifyAiProviderAccess.ts`):
+
+1. Read the id param, `Number.parseInt`/validate it.
+2. Load the row by id.
+3. `404` if it doesn't exist.
+4. Resolve the row's `orgId`, then check/attach `req.userOrg` (query
+   `userOrgs` if not already on the request), `403` if the user isn't in
+   that org.
+5. Run `checkOrgAccessPolicy` if `req.orgPolicyAllowed` hasn't been resolved
+   yet.
+6. Set `req.userOrgId`, `req.userOrgRoleIds`, and stash the row on the
+   request (e.g. `req.aiProvider = provider`) so downstream handlers and
+   `verifyUserHasAction` don't have to refetch it.
+
+Org-scoped create/list routes (`/org/:orgId/x`) don't need a bespoke
+middleware — they use the existing generic `verifyOrgAccess` from
+`@server/middlewares`.
+
+## 5. Register an action + permission check
+
+Add one `ActionsEnum` entry per operation in `server/auth/actions.ts`,
+grouped near the entity's other actions, named `create<Entity>`,
+`get<Entity>`, `update<Entity>`, `delete<Entity>`, `list<Entities>`:
+
+```ts
+createAiProvider = "createAiProvider",
+deleteAiProvider = "deleteAiProvider",
+getAiProvider = "getAiProvider",
+listAiProviders = "listAiProviders",
+updateAiProvider = "updateAiProvider",
+```
+
+Every route uses `verifyUserHasAction(ActionsEnum.x)` to check the caller's
+role/permissions for that action, and mutating routes (create/update/delete)
+follow it with `logActionAudit(ActionsEnum.x)` to record the action in the
+audit log.
+
+## 6. Register the routes
+
+There are four router files; which one(s) you touch depends on public vs.
+private and user-facing vs. service-to-service:
+
+| File | Purpose |
+|---|---|
+| `server/routers/external.ts` | Public, user-facing API. Exports `authenticated`, `unauthenticated`, `authRouter` Express routers. |
+| `server/routers/internal.ts` | Public, internal service-to-service API (gerbil, badger, traefik-config) — no user auth, exports `internalRouter`. |
+| `server/private/routers/external.ts` | Enterprise-only, user-facing. Imports `authenticated`/`unauthenticated`/`authRouter` **from the public `external.ts`** and re-exports them, then adds more routes on top. |
+| `server/private/routers/internal.ts` | Enterprise-only, service-to-service. Same re-export trick with `internalRouter`. |
+
+Private router files always start:
+
+```ts
+import {
+    unauthenticated as ua,
+    authenticated as a,
+    authRouter as aa
+} from "@server/routers/external";
+
+export const authenticated = a;
+export const unauthenticated = ua;
+export const authRouter = aa;
+```
+
+...and then call `authenticated.get/put/post/delete(...)` to bolt on
+additional, enterprise-only routes on the *same* router instances the public
+build uses. This is why the private build has strictly more routes than the
+public build, not a divergent copy.
+
+### Route registration order (mutating vs read)
+
+Standard middleware chain per verb, using `alertRule`'s registrations as the
+template:
+
+```ts
+// Create — org-scoped, no row exists yet
+authenticated.put(
+    "/org/:orgId/x",
+    verifyValidLicense,              // private/enterprise routes only
+    verifyOrgAccess,
+    verifyLimits,                    // if the entity counts against a plan limit
+    verifyUserHasAction(ActionsEnum.createX),
+    logActionAudit(ActionsEnum.createX),
+    x.createX
+);
+
+// Update — row-scoped
+authenticated.post(
+    "/org/:orgId/x/:xId",            // or "/x/:xId" if id is globally unique
+    verifyValidLicense,
+    verifyOrgAccess,                 // or verifyXAccess if globally-keyed
+    verifyUserHasAction(ActionsEnum.updateX),
+    logActionAudit(ActionsEnum.updateX),
+    x.updateX
+);
+
+// Delete — row-scoped
+authenticated.delete(
+    "/org/:orgId/x/:xId",
+    verifyValidLicense,
+    verifyOrgAccess,
+    verifyUserHasAction(ActionsEnum.deleteX),
+    logActionAudit(ActionsEnum.deleteX),
+    x.deleteX
+);
+
+// List — org-scoped, read-only, no audit log
+authenticated.get(
+    "/org/:orgId/xs",
+    verifyValidLicense,
+    verifyOrgAccess,
+    verifyUserHasAction(ActionsEnum.listXs),
+    x.listXs
+);
+
+// Get one — row-scoped, read-only, no audit log
+authenticated.get(
+    "/org/:orgId/x/:xId",
+    verifyValidLicense,
+    verifyOrgAccess,
+    verifyUserHasAction(ActionsEnum.getX),
+    x.getX
+);
+```
+
+Notes:
+
+- HTTP verbs: `PUT` = create, `POST` = update, `GET` = read, `DELETE` =
+  delete. This repo does not use `PATCH` for entity updates (site
+  provisioning keys are the one exception, using `PATCH`).
+- `verifyValidLicense` is only needed on private/enterprise routes; public
+  OSS routes skip it.
+- Use `verifyValidSubscription(tierMatrix.someFeature)` right after
+  `verifyValidLicense` when a feature is gated to specific SaaS tiers (see
+  `tierMatrix` usages in `server/private/routers/external.ts`).
+- `verifyLimits` goes on create routes for entities that count against a
+  plan/seat limit.
+- For entities keyed by a globally-unique id (not nested under `/org/:orgId`),
+  use the dedicated `verify<Entity>Access` middleware from §4 instead of
+  `verifyOrgAccess` on the row-scoped routes (see how `/ai-provider/:providerId`
+  uses `verifyAiProviderAccess`, while `/org/:orgId/ai-provider` create/list
+  use plain `verifyOrgAccess`).
+- Read-only routes (`get`, `list`) skip `logActionAudit` — only mutations are
+  audited.
+- `internal*.ts` routes are for trusted internal callers (gerbil/badger
+  sidecars) and generally skip user-facing auth entirely, using
+  `verifySessionUserMiddleware` / `verifyUserFromResourceSessionMiddleware`
+  instead of `verifyOrgAccess`/`verifyUserHasAction`. CRUD entities almost
+  never need internal router entries — only add one if a sidecar process
+  needs direct access to the resource.
+
+## 7. The `#dynamic` alias (advanced — most CRUD work can ignore this)
+
+Some middleware (e.g. `logActionAudit`) needs a real implementation in the
+enterprise/SaaS build but a no-op stub in the open-source build, while
+being imported by identical code in `server/routers/external.ts` in both
+builds. That's done via the `#dynamic/*` import alias, which
+`tsconfig.oss.json` points at `./server/*` and `tsconfig.enterprise.json` /
+`tsconfig.saas.json` point at `./server/private/*`. You only need this
+pattern if you're adding a genuinely dual-implementation hook; a normal
+private-only CRUD entity (like `alertRule`) never touches `#dynamic` — it
+just lives entirely under `server/private/` and is imported with `#private/*`
+directly from `server/private/routers/external.ts`.
+
+## 8. Checklist for a new entity
+
+1. Add the DB table to `server/db/pg/schema/schema.ts` (and sqlite schema if
+   applicable).
+2. Add `ActionsEnum` entries in `server/auth/actions.ts`.
+3. Create `server/routers/<entity>/` (or `server/private/routers/<entity>/`):
+   `types.ts`, optional `validation.ts`, one file per operation, `index.ts`
+   barrel.
+4. If routes are row-scoped by a global id, add
+   `verify<Entity>Access.ts` to `server/middlewares/` or
+   `server/private/middlewares/`, and export it from that directory's
+   `index.ts`.
+5. Wire routes into `external.ts` (public or private) following the verb/
+   middleware table in §6. Add to `internal.ts` only if a sidecar needs
+   direct access.
+6. Add license header block to every new file if it's under `server/private/`.

messages/en-US.json

@@ -1079,6 +1079,8 @@
     "accessRoleErrorNewRequired": "New role is required",
     "accessRoleErrorRemove": "Failed to remove role",
     "accessRoleErrorRemoveDescription": "An error occurred while removing the role.",
+    "accessRoleInferenceBudget": "Inference Budget",
+    "accessRoleInferenceBudgetDescription": "Configure how members of this role restrict AI usage based on spending or token limits",
     "accessRoleName": "Role Name",
     "accessRoleQuestionRemove": "You're about to delete the `{name}` role. You cannot undo this action.",
     "accessRoleRemove": "Remove Role",
@@ -1424,6 +1426,16 @@
     "actionDeleteSite": "Delete Site",
     "actionGetSite": "Get Site",
     "actionListSites": "List Sites",
+    "actionCreateAiProvider": "Create AI Provider",
+    "actionDeleteAiProvider": "Delete AI Provider",
+    "actionGetAiProvider": "Get AI Provider",
+    "actionListAiProviders": "List AI Providers",
+    "actionUpdateAiProvider": "Update AI Provider",
+    "actionCreateAiModel": "Create AI Model",
+    "actionDeleteAiModel": "Delete AI Model",
+    "actionGetAiModel": "Get AI Model",
+    "actionListAiModels": "List AI Models",
+    "actionUpdateAiModel": "Update AI Model",
     "actionApplyBlueprint": "Apply Blueprint",
     "actionListBlueprints": "List Blueprints",
     "actionGetBlueprint": "Get Blueprint",
@@ -1622,6 +1634,231 @@
     "sidebarInvitations": "Invitations",
     "sidebarRoles": "Roles",
     "sidebarShareableLinks": "Shareable Links",
+    "sidebarAiGateway": "AI Gateway",
+    "sidebarAiProviders": "Providers",
+    "commandAiProviders": "AI Providers",
+    "aiProvidersTitle": "AI Providers",
+    "aiProvidersDescription": "Connect model providers for AI workloads in this organization",
+    "aiProvidersAdd": "Add Provider",
+    "aiProvidersSearch": "Search providers...",
+    "aiProvidersEmpty": "No AI providers yet",
+    "aiProviderCreate": "Create AI Provider",
+    "aiProviderCreateDescription": "Add a model provider for this organization",
+    "aiProviderSeeAll": "See All Providers",
+    "aiProviderSetting": "Provider Settings for {providerName}",
+    "aiProviderSettingDescription": "Configure this AI provider",
+    "aiProviderGeneral": "General",
+    "aiProviderGeneralDescription": "Basic settings for this provider",
+    "aiProviderConfiguration": "Configuration",
+    "aiProviderConfigurationDescription": "Network routing and authentication for this provider",
+    "aiProviderNetworkSettings": "Network Settings",
+    "aiProviderNetworkSettingsDescription": "Choose how traffic reaches this provider",
+    "aiProviderAuthSettings": "Authentication",
+    "aiProviderAuthSettingsDescription": "Configure how this provider authenticates requests to its upstream URL",
+    "aiProviderBudgetSettings": "Budget",
+    "aiProviderBudgetSettingsDescription": "Configure how this provider restricts usage based on spending or token limits",
+    "aiBudgetAdd": "Add Budget",
+    "aiBudgetEmpty": "No budgets configured yet. Click Add Budget to set a spending or token limit.",
+    "aiBudgetUnit": "Spend Type",
+    "aiBudgetPeriod": "Reset Period",
+    "aiBudgetAmount": "Maximum Spend",
+    "aiBudgetAmountPlaceholder": "Maximum spend",
+    "aiBudgetPeriodHourly": "Hourly",
+    "aiBudgetPeriodDaily": "Daily",
+    "aiBudgetPeriodWeekly": "Weekly",
+    "aiBudgetPeriodMonthly": "Monthly",
+    "aiBudgetPeriodYearly": "Yearly",
+    "aiBudgetPeriodLifetime": "Lifetime",
+    "aiBudgetUnitUsd": "USD",
+    "aiBudgetUnitTokens": "Tokens",
+    "aiBudgetConflictError": "A budget for this reset period and spend type already exists",
+    "aiBudgetInvalidAmountError": "Enter a maximum spend greater than 0",
+    "aiBudgetUpdated": "Budgets updated",
+    "aiBudgetErrorSave": "Failed to update budgets",
+    "aiProviderType": "Provider Type",
+    "aiProviderTypeSearch": "Search providers...",
+    "aiProviderTypeNotFound": "No provider type found",
+    "aiProviderTypeOpenai": "OpenAI",
+    "aiProviderTypeAnthropic": "Anthropic",
+    "aiProviderTypeGoogleGemini": "Google Gemini",
+    "aiProviderTypeVertexAi": "Vertex AI",
+    "aiProviderTypeBedrock": "Amazon Bedrock",
+    "aiProviderTypeMicrosoftFoundry": "Microsoft Foundry",
+    "aiProviderTypeOpenRouter": "OpenRouter",
+    "aiProviderTypeVercelAiGateway": "Vercel AI Gateway",
+    "aiProviderTypeCustom": "Custom",
+    "aiProviderTypeOpenaiDescription": "OpenAI API with default upstream URL",
+    "aiProviderTypeAnthropicDescription": "Anthropic API with default upstream URL",
+    "aiProviderTypeGoogleGeminiDescription": "Google Gemini generateContent API",
+    "aiProviderTypeVertexAiDescription": "Google Vertex AI; upstream URL required",
+    "aiProviderTypeBedrockDescription": "Amazon Bedrock Runtime",
+    "aiProviderTypeMicrosoftFoundryDescription": "Microsoft Foundry; upstream URL required",
+    "aiProviderTypeOpenRouterDescription": "OpenRouter API",
+    "aiProviderTypeVercelAiGatewayDescription": "Vercel AI Gateway",
+    "aiProviderTypeCustomDescription": "Bring your own OpenAI-compatible endpoint or route via Pangolin targets",
+    "aiProviderUpstreamUrl": "Upstream URL",
+    "aiProviderUpstreamUrlDescription": "Base URL for the provider API",
+    "aiProviderUpstreamUrlOptionalDescription": "Leave blank to use the default upstream URL for this provider",
+    "aiProviderEffectiveUpstreamUrl": "Effective Upstream URL",
+    "aiProviderApiKey": "API Key",
+    "aiProviderApiKeyDescription": "API key used to authenticate requests to this provider",
+    "aiProviderCustomHeadersDescription": "Headers sent on every request to this provider. Newline separated: Header-Name: value",
+    "aiProviderApiKeyLastChars": "API Key",
+    "aiProviderAuthType": "Auth Type",
+    "aiProviderAuthTypeSearch": "Search auth types...",
+    "aiProviderAuthTypeNotFound": "No auth type found",
+    "aiProviderAuthTypeBearer": "Bearer",
+    "aiProviderAuthTypeBearerDescription": "Authorization: Bearer key. Used by OpenAI and most providers",
+    "aiProviderAuthTypeXApiKey": "x-api-key",
+    "aiProviderAuthTypeXApiKeyDescription": "x-api-key header. Used by Anthropic",
+    "aiProviderAuthTypeXGoogApiKey": "x-goog-api-key",
+    "aiProviderAuthTypeXGoogApiKeyDescription": "x-goog-api-key header. Used by Google Gemini",
+    "aiProviderAuthTypeHec": "Splunk HEC",
+    "aiProviderAuthTypeHecDescription": "Authorization: Splunk key. Used by Splunk HTTP Event Collector",
+    "aiProviderAuthTypeCfAigAuthorization": "Cloudflare AI Gateway",
+    "aiProviderAuthTypeCfAigAuthorizationDescription": "cf-aig-authorization: Bearer key. Used by Cloudflare AI Gateway",
+    "aiProviderAuthTypeNone": "No Auth",
+    "aiProviderAuthTypePassthrough": "Passthrough",
+    "aiProviderAuthTypeDescription": "How the upstream API authenticates requests",
+    "aiProviderAuthTypePassthroughDescription": "Forward the caller's API key headers to the upstream",
+    "aiProviderAuthTypeNoneDescription": "Do not send authentication headers to the upstream",
+    "aiProviderRoutingMode": "Routing Mode",
+    "aiProviderRoutingModeDescription": "Send traffic to an upstream URL or to HTTP targets on your sites",
+    "aiProviderRoutingModeUrl": "Upstream URL",
+    "aiProviderRoutingModeUrlDescription": "Call a public or private API base URL",
+    "aiProviderRoutingModeTarget": "Site Targets",
+    "aiProviderRoutingModeTargetDescription": "Route through targets on your sites",
+    "aiProviderRoutingModeTargetNote": "After creating this provider, configure site targets on the Network Settings tab.",
+    "aiProviderTargetNoOne": "This provider doesn't have any targets. Add a target to route requests through your sites.",
+    "aiProviderSkipTlsVerification": "Skip TLS Verification",
+    "aiProviderSkipTlsVerificationDescription": "Disable TLS certificate verification for the upstream connection",
+    "aiProviderBudget": "Budget",
+    "aiProviderBudgetDescription": "Optional spending or token budget for this provider",
+    "aiProviderBudgetAmount": "Budget Amount",
+    "aiProviderBudgetUnit": "Budget Unit",
+    "aiProviderBudgetUnitUsd": "USD",
+    "aiProviderBudgetUnitTokens": "Tokens",
+    "aiProviderEnabled": "Enabled",
+    "aiProviderEnabledDescription": "Disable to stop using this provider without deleting it",
+    "aiProviderErrorCreate": "Failed to create AI provider",
+    "aiProviderErrorUpdate": "Failed to update AI provider",
+    "aiProviderErrorDelete": "Failed to delete AI provider",
+    "aiProviderErrorLoad": "Failed to load AI provider",
+    "aiProviderErrorUpstreamUrlInvalid": "Enter a valid upstream URL",
+    "aiProviderErrorUpstreamUrlRequired": "Upstream URL is required for this provider",
+    "aiProviderErrorAuthTypeRequired": "Auth type is required",
+    "aiProviderErrorApiKeyRequired": "API key is required",
+    "aiProviderErrorRoutingModeTarget": "Site targets routing is only available for custom providers",
+    "aiProviderErrorCapabilitiesRequired": "Select at least one API capability",
+    "aiProviderCapabilities": "API Capabilities",
+    "aiProviderCapabilitiesDescription": "Select which API formats this provider can handle. Known providers start with recommended defaults.",
+    "aiProviderCapabilitiesCustomDescription": "Select which API formats this custom provider can handle",
+    "aiProviderCapabilitiesSelect": "Select capabilities",
+    "aiProviderCapabilitiesEmpty": "No capabilities found",
+    "aiProviderCapabilitiesSearch": "Search capabilities...",
+    "aiCapabilityOpenaiChat": "OpenAI Chat Completions",
+    "aiCapabilityOpenaiChatDescription": "Supports /v1/chat/completions",
+    "aiCapabilityOpenaiResponses": "OpenAI Responses",
+    "aiCapabilityOpenaiResponsesDescription": "Supports /v1/responses",
+    "aiCapabilityAnthropicMessages": "Anthropic Messages",
+    "aiCapabilityAnthropicMessagesDescription": "Supports /v1/messages",
+    "aiCapabilityGeminiGenerateContent": "Gemini Generate Content",
+    "aiCapabilityGeminiGenerateContentDescription": "Supports the direct Gemini API",
+    "aiCapabilityBedrockModelInvoke": "Bedrock Model Invoke",
+    "aiCapabilityBedrockModelInvokeDescription": "Supports Amazon Bedrock InvokeModel",
+    "aiCapabilityGoogleGenerateContent": "Vertex Generate Content",
+    "aiCapabilityGoogleGenerateContentDescription": "Supports Vertex AI Gemini format",
+    "aiCapabilityGoogleRawPredict": "Vertex Raw Predict",
+    "aiCapabilityGoogleRawPredictDescription": "Supports Vertex AI rawPredict for Anthropic models",
+    "aiCapabilityBedrockConverse": "Bedrock Converse",
+    "aiCapabilityBedrockConverseDescription": "Supports Amazon Bedrock Converse API",
+    "aiProviderCreated": "AI provider created",
+    "aiProviderUpdated": "AI provider updated",
+    "aiProviderDeleted": "AI provider deleted",
+    "aiProviderDelete": "Delete Provider",
+    "aiProviderDeleteConfirm": "Delete Provider",
+    "aiProviderQuestionRemove": "Are you sure you want to delete this AI provider?",
+    "aiProviderMessageRemove": "This will permanently delete the provider and its models and targets. This cannot be undone.",
+    "aiProviderErrorNoUpdate": "AI provider is not available to update",
+    "aiProviderModels": "Models",
+    "aiProviderModelsDescription": "Define allow and block lists for this provider. Requests must match an allow entry and must not match a block entry.",
+    "aiProviderModelsPlaceholder": "Search models or type a custom key",
+    "aiProviderModelsAllow": "Allow List",
+    "aiProviderModelsAllowDescription": "Models that may be used through this provider. Empty means deny all.",
+    "aiProviderModelsAllowPlaceholder": "Enter model key",
+    "aiProviderModelsAllowEmpty": "No models allowed. All requests will be denied.",
+    "aiProviderModelsBlock": "Block List",
+    "aiProviderModelsBlockDescription": "Models to deny even if they match an allow entry.",
+    "aiProviderModelsBlockPlaceholder": "Enter model key",
+    "aiProviderModelsBlockEmpty": "No blocked models.",
+    "aiProviderModelsAdd": "Add Models",
+    "aiProviderModelsClearAll": "Clear All",
+    "aiProviderModelsAddCustom": "Add \"{key}\"",
+    "aiProviderModelsAddCustomHint": "Press Enter to add this custom model key.",
+    "aiProviderModelsAddBulk": "Add {count} custom keys",
+    "aiProviderModelsAddBulkHint": "Press Enter to add {count} custom keys.",
+    "aiProviderModelsAddOne": "Add {key} now",
+    "aiProviderModelsAddSelected": "Add Selected",
+    "aiProviderModelsSelectedCount": "{count} selected",
+    "aiProviderModelsSelectAll": "Select all",
+    "aiProviderModelsClearSelected": "Clear",
+    "aiProviderModelsBulkHint": "Select known models or type a custom key.",
+    "aiProviderModelsCatalogEmpty": "No matching catalog models.",
+    "aiProviderModelsCatalogHeading": "Known Models",
+    "aiProviderModelsAllLabel": "All models",
+    "aiProviderModelsAllPatternHint": "Wildcard: *",
+    "aiProviderModelsAddAllAllow": "Allow all models",
+    "aiProviderModelsAddAllBlock": "Block all models",
+    "aiProviderModelsAddAllDescription": "Uses the * wildcard so every model key matches.",
+    "aiProviderModelsViewMore": "View more ({count})",
+    "aiProviderModelsViewLess": "View less",
+    "aiProviderModelsRemove": "Remove model",
+    "aiProviderModelsEditHint": "Click to edit model settings",
+    "aiProviderModelsSourceCatalog": "Known catalog model",
+    "aiProviderModelsSourceCustom": "Custom model key",
+    "aiProviderModelsSourcePattern": "Wildcard pattern",
+    "aiProviderModelsSourceAll": "Matches every model key",
+    "aiProviderModelsBudgetConfigured": "Budget configured",
+    "aiProviderModelsEditTitle": "Edit Model",
+    "aiProviderModelsEditDescription": "Update the model key or configure its inference budget.",
+    "aiProviderModelsBudgetTab": "Inference Budget",
+    "aiProviderModelsBudgetDescription": "Configure how this model restricts usage based on spending or token limits",
+    "aiProviderModelsBudgetUnsaved": "Save this model first to configure its inference budget.",
+    "aiProviderModelsKeyLabel": "Model Key",
+    "aiProviderModelsKeyRequired": "Enter a model key",
+    "aiProviderModelsKeyDuplicate": "This model key is already on a list",
+    "aiProviderModelsOverlapError": "These patterns cannot be on both lists: {keys}",
+    "aiProviderModelsUpdated": "Models updated",
+    "aiProviderModelsErrorUpdate": "Failed to update models",
+    "aiResourceProviders": "Providers",
+    "aiResourceProvidersDescription": "Choose which AI providers this inference resource can use",
+    "aiResourceProvidersHelp": "Attach providers and choose inherit (use each provider's lists) or select (pick an allow list for this resource). Allow patterns that conflict across attached providers are not allowed.",
+    "aiResourceProvidersSelect": "Select providers",
+    "aiResourceProvidersEmpty": "No AI providers found",
+    "aiResourceProvidersNoneAttached": "No providers attached yet.",
+    "aiResourceProvidersAdd": "Add provider",
+    "aiResourceProvidersRemove": "Remove provider",
+    "aiResourceProviderToggleEnabled": "Enable or disable this provider on the resource",
+    "aiResourceProviderDisabled": "Disabled",
+    "aiResourceProvidersUpdated": "Providers updated",
+    "aiResourceProvidersErrorUpdate": "Failed to update providers",
+    "aiResourceProviderEditDescription": "Choose how this provider's models are exposed on this resource.",
+    "aiResourceProviderMode": "Access mode",
+    "aiResourceProviderModeInherit": "Inherit",
+    "aiResourceProviderModeSelect": "Select",
+    "aiResourceProviderModeSelectSummary": "Select · {count} models",
+    "aiResourceProviderModeInheritHelp": "Use this provider's allow and block lists as configured on the provider.",
+    "aiResourceProviderModeSelectHelp": "Choose a subset of this provider's allow-list models for this resource.",
+    "aiResourceProviderAllowModels": "Allow list",
+    "aiResourceProviderAllowModelsSelect": "Select models",
+    "aiResourceProviderAllowModelsSearch": "Search models...",
+    "aiResourceProviderAllowModelsEmpty": "No models found",
+    "aiResourceProviderAllowModelsHelp": "Only models from this provider's allow list can be selected.",
+    "aiResourceAliasRequired": "Alias is required for inference resources",
+    "aiResourceDomainConfiguration": "Domain configuration",
+    "aiResourceDomainConfigurationDescription": "Choose the domain clients will use to reach this inference resource.",
+    "resourceBudgetSettings": "Budget",
+    "resourceBudgetSettingsDescription": "Configure how this inference resource restricts usage based on spending or token limits",
     "sidebarApiKeys": "API Keys",
     "sidebarProvisioning": "Provisioning",
     "sidebarSettings": "Settings",
@@ -2246,6 +2483,7 @@
     "requireDeviceApproval": "Require Device Approvals",
     "requireDeviceApprovalDescription": "Users with this role need new devices approved by an admin before they can connect and access resources.",
     "sshSettings": "SSH Settings",
+    "inferenceSettings": "Inference Settings",
     "sshAccess": "SSH Access",
     "rdpSettings": "RDP Settings",
     "vncSettings": "VNC Settings",
@@ -2401,6 +2639,7 @@
     "editInternalResourceDialogModeCidr": "CIDR",
     "editInternalResourceDialogModeHttp": "HTTP",
     "editInternalResourceDialogModeHttps": "HTTPS",
+    "editInternalResourceDialogModeInference": "Inference",
     "editInternalResourceDialogModeSsh": "SSH",
     "editInternalResourceDialogScheme": "Scheme",
     "editInternalResourceDialogEnableSsl": "Enable TLS",
@@ -2460,6 +2699,7 @@
     "createInternalResourceDialogModeHttp": "HTTP",
     "createInternalResourceDialogModeHttps": "HTTPS",
     "createInternalResourceDialogModeSsh": "SSH",
+    "createInternalResourceDialogModeInference": "Inference",
     "scheme": "Scheme",
     "createInternalResourceDialogScheme": "Scheme",
     "createInternalResourceDialogEnableSsl": "Enable TLS",
@@ -3826,4 +4066,4 @@
     "sessionToolbarShow": "Show toolbar",
     "sessionToolbarHide": "Hide toolbar",
     "actionUpdateSiteApprovals": "Update Site Approvals"
-}
+}

package.json

@@ -32,49 +32,49 @@
         "format": "prettier --write ."
     },
     "dependencies": {
-        "@asteasolutions/zod-to-openapi": "9.1.0",
+        "@asteasolutions/zod-to-openapi": "8.5.0",
+        "@aws-sdk/client-s3": "3.1056.0",
         "@devolutions/iron-remote-desktop": "https://static.pangolin.net/packages/devolutions-iron-remote-desktop-0.0.0.tgz",
         "@devolutions/iron-remote-desktop-rdp": "https://static.pangolin.net/packages/devolutions-iron-remote-desktop-rdp-0.0.1.tgz",
-        "@aws-sdk/client-s3": "3.1105.0",
         "@headlessui/react": "2.2.10",
-        "@hookform/resolvers": "5.7.1",
+        "@hookform/resolvers": "5.4.0",
         "@monaco-editor/react": "4.7.0",
         "@node-rs/argon2": "2.0.2",
         "@novnc/novnc": "^1.7.0",
         "@oslojs/crypto": "1.0.1",
         "@oslojs/encoding": "1.1.0",
-        "@radix-ui/react-avatar": "1.2.6",
-        "@radix-ui/react-checkbox": "1.3.11",
-        "@radix-ui/react-collapsible": "1.1.20",
-        "@radix-ui/react-dialog": "1.1.23",
-        "@radix-ui/react-dropdown-menu": "2.1.24",
+        "@radix-ui/react-avatar": "1.1.11",
+        "@radix-ui/react-checkbox": "1.3.3",
+        "@radix-ui/react-collapsible": "1.1.12",
+        "@radix-ui/react-dialog": "1.1.15",
+        "@radix-ui/react-dropdown-menu": "2.1.16",
         "@radix-ui/react-icons": "1.3.2",
-        "@radix-ui/react-label": "2.1.15",
-        "@radix-ui/react-popover": "1.1.23",
-        "@radix-ui/react-progress": "1.1.16",
-        "@radix-ui/react-radio-group": "1.4.7",
-        "@radix-ui/react-scroll-area": "1.2.18",
-        "@radix-ui/react-select": "2.3.7",
-        "@radix-ui/react-separator": "1.1.15",
-        "@radix-ui/react-slot": "1.3.3",
-        "@radix-ui/react-switch": "1.3.7",
-        "@radix-ui/react-tabs": "1.1.21",
-        "@radix-ui/react-toast": "1.2.23",
-        "@radix-ui/react-tooltip": "1.2.16",
+        "@radix-ui/react-label": "2.1.8",
+        "@radix-ui/react-popover": "1.1.15",
+        "@radix-ui/react-progress": "1.1.8",
+        "@radix-ui/react-radio-group": "1.3.8",
+        "@radix-ui/react-scroll-area": "1.2.10",
+        "@radix-ui/react-select": "2.2.6",
+        "@radix-ui/react-separator": "1.1.8",
+        "@radix-ui/react-slot": "1.2.4",
+        "@radix-ui/react-switch": "1.2.6",
+        "@radix-ui/react-tabs": "1.1.13",
+        "@radix-ui/react-toast": "1.2.15",
+        "@radix-ui/react-tooltip": "1.2.8",
         "@react-email/body": "0.3.0",
         "@react-email/components": "1.0.12",
-        "@react-email/render": "2.1.0",
+        "@react-email/render": "2.0.8",
         "@react-email/tailwind": "2.0.7",
         "@simplewebauthn/browser": "13.3.0",
-        "@simplewebauthn/server": "13.3.2",
+        "@simplewebauthn/server": "13.3.1",
         "@tailwindcss/forms": "0.5.11",
-        "@tanstack/react-query": "5.101.4",
-        "@tanstack/react-table": "9.0.0",
+        "@tanstack/react-query": "5.100.14",
+        "@tanstack/react-table": "8.21.3",
         "@xterm/addon-fit": "^0.11.0",
         "@xterm/addon-web-links": "^0.12.0",
         "@xterm/xterm": "^6.0.0",
         "arctic": "3.7.0",
-        "axios": "1.19.0",
+        "axios": "1.18.0",
         "better-sqlite3": "11.9.1",
         "canvas-confetti": "1.9.4",
         "class-variance-authority": "0.7.1",
@@ -86,61 +86,62 @@
         "d3": "7.9.0",
         "drizzle-orm": "0.45.2",
         "express": "5.2.1",
-        "express-rate-limit": "8.6.2",
+        "express-rate-limit": "8.5.2",
         "glob": "13.0.6",
-        "helmet": "8.3.0",
+        "gpt-tokenizer": "^3.4.0",
+        "helmet": "8.2.0",
         "http-errors": "2.0.1",
         "input-otp": "1.4.2",
-        "ioredis": "6.0.0",
+        "ioredis": "5.11.0",
         "jmespath": "0.16.0",
-        "js-yaml": "5.2.3",
+        "js-yaml": "4.3.0",
         "jsonwebtoken": "9.0.3",
-        "lucide-react": "1.29.0",
-        "maxmind": "5.0.7",
+        "lucide-react": "1.17.0",
+        "maxmind": "5.0.6",
         "moment": "2.30.1",
-        "next": "16.3.0",
-        "next-intl": "4.13.5",
+        "next": "16.2.11",
+        "next-intl": "4.13.0",
         "next-themes": "0.4.6",
         "nextjs-toploader": "3.9.17",
         "node-cache": "5.1.2",
-        "nodemailer": "9.0.4",
+        "nodemailer": "9.0.1",
         "oslo": "1.2.1",
-        "pg": "8.22.0",
-        "posthog-node": "5.48.1",
+        "pg": "8.21.0",
+        "posthog-node": "5.35.6",
         "qrcode.react": "4.2.0",
-        "react": "19.2.8",
-        "react-day-picker": "10.0.1",
-        "react-dom": "19.2.8",
+        "react": "19.2.6",
+        "react-day-picker": "9.14.0",
+        "react-dom": "19.2.6",
         "react-easy-sort": "1.8.0",
-        "react-hook-form": "7.84.0",
-        "react-icons": "5.7.0",
-        "recharts": "3.10.1",
+        "react-hook-form": "7.76.1",
+        "react-icons": "5.6.0",
+        "recharts": "3.8.1",
         "reodotdev": "1.1.0",
-        "semver": "7.8.5",
+        "semver": "7.8.1",
         "sshpk": "1.18.0",
-        "stripe": "22.4.0",
+        "stripe": "22.2.0",
         "swagger-ui-express": "5.0.1",
         "tailwind-merge": "3.6.0",
         "topojson-client": "3.1.0",
         "tw-animate-css": "1.4.0",
         "use-debounce": "10.1.1",
-        "uuid": "14.0.1",
+        "uuid": "14.0.0",
         "vaul": "1.1.2",
         "visionscarto-world-atlas": "1.0.0",
         "winston": "3.19.0",
         "winston-daily-rotate-file": "5.0.0",
-        "ws": "8.21.2",
+        "ws": "8.21.0",
         "yaml": "2.9.0",
-        "yargs": "18.1.0",
+        "yargs": "18.0.0",
         "zod": "4.4.3",
         "zod-validation-error": "5.0.0"
     },
     "devDependencies": {
-        "@dotenvx/dotenvx": "2.19.2",
+        "@dotenvx/dotenvx": "1.69.1",
         "@esbuild-plugins/tsconfig-paths": "0.1.2",
-        "@react-email/ui": "^6.9.1",
-        "@tailwindcss/postcss": "4.3.3",
-        "@tanstack/react-query-devtools": "5.101.4",
+        "@react-email/ui": "^6.5.0",
+        "@tailwindcss/postcss": "4.3.0",
+        "@tanstack/react-query-devtools": "5.100.14",
         "@types/better-sqlite3": "7.6.13",
         "@types/cookie-parser": "1.4.10",
         "@types/cors": "2.8.19",
@@ -151,14 +152,14 @@
         "@types/jmespath": "0.15.2",
         "@types/js-yaml": "4.0.9",
         "@types/jsonwebtoken": "9.0.10",
-        "@types/node": "26.1.2",
-        "@types/nodemailer": "8.0.1",
+        "@types/node": "25.9.1",
+        "@types/nodemailer": "8.0.0",
         "@types/nprogress": "0.2.3",
-        "@types/pg": "8.20.4",
-        "@types/react": "19.2.18",
-        "@types/react-dom": "19.2.4",
-        "@types/semver": "7.8.0",
-        "@types/sshpk": "1.17.5",
+        "@types/pg": "8.20.0",
+        "@types/react": "19.2.15",
+        "@types/react-dom": "19.2.3",
+        "@types/semver": "7.7.1",
+        "@types/sshpk": "1.17.4",
         "@types/swagger-ui-express": "4.1.8",
         "@types/topojson-client": "3.1.5",
         "@types/ws": "8.18.1",
@@ -166,21 +167,21 @@
         "babel-plugin-react-compiler": "1.0.0",
         "drizzle-kit": "0.31.10",
         "esbuild": "0.28.1",
-        "esbuild-node-externals": "2.0.0",
-        "eslint": "10.8.0",
-        "eslint-config-next": "16.3.0",
-        "postcss": "8.5.26",
-        "prettier": "3.9.6",
-        "react-email": "6.9.1",
-        "tailwindcss": "4.3.3",
-        "tsc-alias": "1.9.1",
-        "tsx": "4.23.9",
-        "typescript": "7.0.2",
-        "typescript-eslint": "8.66.0"
+        "esbuild-node-externals": "1.22.0",
+        "eslint": "10.4.0",
+        "eslint-config-next": "16.2.6",
+        "postcss": "8.5.15",
+        "prettier": "3.8.3",
+        "react-email": "6.5.0",
+        "tailwindcss": "4.3.0",
+        "tsc-alias": "1.8.17",
+        "tsx": "4.22.3",
+        "typescript": "6.0.3",
+        "typescript-eslint": "8.60.0"
     },
     "overrides": {
         "esbuild": "0.28.1",
         "dompurify": "3.4.0",
-        "postcss": "8.5.26"
+        "postcss": "8.5.15"
     }
 }

server/aiGatewayServer.ts

@@ -0,0 +1,39 @@
+import express from "express";
+import helmet from "helmet";
+import cors from "cors";
+import config from "@server/lib/config";
+import logger from "@server/logger";
+import {
+    errorHandlerMiddleware,
+    notFoundMiddleware
+} from "@server/middlewares";
+import { createAiGatewayRouter } from "@server/routers/aiGateway";
+
+const aiGatewayPort = config.getRawConfig().server.ai_gateway_port;
+
+export function createAiGatewayServer() {
+    const aiGatewayServer = express();
+
+    const trustProxy = config.getRawConfig().server.trust_proxy;
+    if (trustProxy) {
+        aiGatewayServer.set("trust proxy", trustProxy);
+    }
+
+    aiGatewayServer.use(helmet());
+    aiGatewayServer.use(cors());
+    aiGatewayServer.use(express.json());
+
+    aiGatewayServer.use(createAiGatewayRouter());
+
+    aiGatewayServer.use(notFoundMiddleware);
+    aiGatewayServer.use(errorHandlerMiddleware);
+
+    aiGatewayServer.listen(aiGatewayPort, (err?: any) => {
+        if (err) throw err;
+        logger.info(
+            `AI gateway server is running on http://localhost:${aiGatewayPort}`
+        );
+    });
+
+    return aiGatewayServer;
+}

server/auth/actions.ts

@@ -50,6 +50,8 @@ export enum ActionsEnum {
     setResourceUsers = "setResourceUsers",
     setResourceRoles = "setResourceRoles",
     listResourceUsers = "listResourceUsers",
+    listResourceAiModels = "listResourceAiModels",
+    setResourceAiModels = "setResourceAiModels",
     // removeRoleSite = "removeRoleSite",
     // addRoleAction = "addRoleAction",
     // removeRoleAction = "removeRoleAction",
@@ -182,7 +184,22 @@ export enum ActionsEnum {
     setResourcePolicyHeaderAuth = "setResourcePolicyHeaderAuth",
     setResourcePolicyWhitelist = "setResourcePolicyWhitelist",
     setResourcePolicyRules = "setResourcePolicyRules",
-    createOrgWideLauncherView = "createOrgWideLauncherView"
+    createOrgWideLauncherView = "createOrgWideLauncherView",
+    createAiProvider = "createAiProvider",
+    deleteAiProvider = "deleteAiProvider",
+    getAiProvider = "getAiProvider",
+    listAiProviders = "listAiProviders",
+    updateAiProvider = "updateAiProvider",
+    createAiModel = "createAiModel",
+    deleteAiModel = "deleteAiModel",
+    getAiModel = "getAiModel",
+    listAiModels = "listAiModels",
+    updateAiModel = "updateAiModel",
+    createAiBudget = "createAiBudget",
+    deleteAiBudget = "deleteAiBudget",
+    getAiBudget = "getAiBudget",
+    listAiBudgets = "listAiBudgets",
+    updateAiBudget = "updateAiBudget"
 }
 
 export async function checkUserActionPermission(

server/db/pg/schema/schema.ts

@@ -3,6 +3,7 @@ import { InferSelectModel, sql } from "drizzle-orm";
 import {
     bigint,
     boolean,
+    check,
     index,
     integer,
     pgTable,
@@ -99,6 +100,7 @@ export const sites = pgTable(
         name: varchar("name").notNull(),
         pubKey: varchar("pubKey"),
         subnet: varchar("subnet"),
+        exitNodeSubnet: text("exitNodeSubnet"), // this is the subnet when connecting to an exit node and INCLUDES THE CIDR
         megabytesIn: real("bytesIn").default(0),
         megabytesOut: real("bytesOut").default(0),
         lastBandwidthUpdate: varchar("lastBandwidthUpdate"),
@@ -194,7 +196,12 @@ export const resources = pgTable(
         postAuthPath: text("postAuthPath"),
         health: varchar("health").default("unknown"), // "healthy", "unhealthy", "unknown"
         wildcard: boolean("wildcard").notNull().default(false),
-        mode: text("mode").default("http").notNull(), // rdp, ssh, http, vnc
+        mode: text("mode")
+            .default("http")
+            .$type<
+                "rdp" | "ssh" | "http" | "vnc" | "inference" | "tcp" | "udp"
+            >()
+            .notNull(),
         pamMode: varchar("pamMode", { length: 32 })
             .$type<"passthrough" | "push">()
             .default("passthrough"),
@@ -215,6 +222,41 @@ export const resources = pgTable(
     ]
 );
 
+export const resourceAiProviders = pgTable(
+    "resourceAiProviders",
+    {
+        resourceId: integer("resourceId")
+            .notNull()
+            .references(() => resources.resourceId, { onDelete: "cascade" }),
+        providerId: integer("providerId")
+            .notNull()
+            .references(() => aiProviders.providerId, { onDelete: "cascade" }),
+        accessMode: varchar("accessMode")
+            .$type<"inherit" | "select">()
+            .notNull()
+            .default("inherit"),
+        enabled: boolean("enabled").notNull().default(true)
+    },
+    (t) => [primaryKey({ columns: [t.resourceId, t.providerId] })]
+);
+
+export const resourceAiModels = pgTable(
+    "resourceAiModels",
+    {
+        resourceId: integer("resourceId")
+            .notNull()
+            .references(() => resources.resourceId, { onDelete: "cascade" }),
+        modelId: integer("modelId")
+            .notNull()
+            .references(() => aiModels.modelId, { onDelete: "cascade" }),
+        listType: varchar("listType")
+            .$type<"allow" | "block">()
+            .notNull()
+            .default("allow")
+    },
+    (t) => [primaryKey({ columns: [t.resourceId, t.modelId] })]
+);
+
 export const labels = pgTable("labels", {
     labelId: serial("labelId").primaryKey(),
     name: varchar("name").notNull(),
@@ -317,11 +359,18 @@ export const targets = pgTable(
     "targets",
     {
         targetId: serial("targetId").primaryKey(),
-        resourceId: integer("resourceId")
-            .references(() => resources.resourceId, {
+        resourceId: integer("resourceId").references(
+            () => resources.resourceId,
+            {
                 onDelete: "cascade"
-            })
-            .notNull(),
+            }
+        ),
+        providerId: integer("providerId").references(
+            () => aiProviders.providerId,
+            {
+                onDelete: "cascade"
+            }
+        ),
         siteId: integer("siteId")
             .references(() => sites.siteId, {
                 onDelete: "cascade"
@@ -345,6 +394,7 @@ export const targets = pgTable(
     },
     (t) => [
         index("idx_targets_resourceid_siteid").on(t.resourceId, t.siteId),
+        index("idx_targets_providerid_siteid").on(t.providerId, t.siteId),
         index("idx_targets_site_enabled_priority_target_resource")
             .on(t.siteId, t.priority.desc(), t.targetId, t.resourceId)
             .where(sql`${t.enabled} = true`)
@@ -424,11 +474,14 @@ export const siteResources = pgTable(
                 onDelete: "restrict"
             }
         ),
+        requiresExitNodeConnection: boolean("requiresExitNodeConnection")
+            .notNull()
+            .default(false),
         niceId: varchar("niceId").notNull(),
         name: varchar("name").notNull(),
         ssl: boolean("ssl").notNull().default(false),
         mode: varchar("mode")
-            .$type<"host" | "cidr" | "http" | "ssh">()
+            .$type<"host" | "cidr" | "http" | "ssh" | "inference">()
             .notNull(), // "host" | "cidr" | "http"
         scheme: varchar("scheme").$type<"http" | "https">(), // only for when we are doing https or http mode
         proxyPort: integer("proxyPort"), // only for port mode
@@ -463,6 +516,45 @@ export const siteResources = pgTable(
     (t) => [index("idx_siteresources_orgid_niceid").on(t.orgId, t.niceId)]
 );
 
+export const siteResourceAiProviders = pgTable(
+    "siteResourceAiProviders",
+    {
+        siteResourceId: integer("siteResourceId")
+            .notNull()
+            .references(() => siteResources.siteResourceId, {
+                onDelete: "cascade"
+            }),
+        providerId: integer("providerId")
+            .notNull()
+            .references(() => aiProviders.providerId, { onDelete: "cascade" }),
+        accessMode: varchar("accessMode")
+            .$type<"inherit" | "select">()
+            .notNull()
+            .default("inherit"),
+        enabled: boolean("enabled").notNull().default(true)
+    },
+    (t) => [primaryKey({ columns: [t.siteResourceId, t.providerId] })]
+);
+
+export const siteResourceAiModels = pgTable(
+    "siteResourceAiModels",
+    {
+        siteResourceId: integer("siteResourceId")
+            .notNull()
+            .references(() => siteResources.siteResourceId, {
+                onDelete: "cascade"
+            }),
+        modelId: integer("modelId")
+            .notNull()
+            .references(() => aiModels.modelId, { onDelete: "cascade" }),
+        listType: varchar("listType")
+            .$type<"allow" | "block">()
+            .notNull()
+            .default("allow")
+    },
+    (t) => [primaryKey({ columns: [t.siteResourceId, t.modelId] })]
+);
+
 export const networks = pgTable(
     "networks",
     {
@@ -1181,7 +1273,7 @@ export const clients = pgTable(
         olmId: text("olmId"), // to lock it to a specific olm optionally
         name: varchar("name").notNull(),
         pubKey: varchar("pubKey"),
-        subnet: varchar("subnet").notNull(),
+        exitNodeSubnet: varchar("exitNodeSubnet").notNull(), // INCLUDES THE CIDR
         megabytesIn: real("bytesIn"),
         megabytesOut: real("bytesOut"),
         lastBandwidthUpdate: varchar("lastBandwidthUpdate"),
@@ -1540,6 +1632,132 @@ export const statusHistory = pgTable(
     ]
 );
 
+export const aiProviders = pgTable("aiProviders", {
+    providerId: serial("providerId").primaryKey(),
+    orgId: varchar("orgId")
+        .notNull()
+        .references(() => orgs.orgId, { onDelete: "cascade" }),
+    name: varchar("name").notNull(),
+    type: varchar("type")
+        .$type<
+            | "openai"
+            | "anthropic"
+            | "googleGemini"
+            | "vertexAi"
+            | "bedrock"
+            | "microsoftFoundry"
+            | "openRouter"
+            | "vercelAiGateway"
+            | "custom"
+        >()
+        .notNull(),
+    upstreamUrl: text("upstreamUrl"),
+    apiKey: text("apiKey"),
+    apiKeyLastChars: varchar("apiKeyLastChars"),
+    authType: varchar("authType")
+        .$type<
+            | "bearer"
+            | "x-api-key"
+            | "x-goog-api-key"
+            | "hec"
+            | "cf-aig-authorization"
+            | "none"
+            | "passthrough"
+        >()
+        .notNull(),
+    routingMode: varchar("routingMode")
+        .$type<"url" | "target">()
+        .notNull()
+        .default("url"),
+    capabilities: text("capabilities").notNull().default("[]"),
+    headers: text("headers"), // JSON array of { name, value }
+    skipTlsVerification: boolean("skipTlsVerification")
+        .notNull()
+        .default(false),
+    enabled: boolean("enabled").notNull().default(true),
+    createdAt: bigint("createdAt", { mode: "number" }).notNull(),
+    updatedAt: bigint("updatedAt", { mode: "number" }).notNull()
+});
+
+export const aiModels = pgTable(
+    "aiModels",
+    {
+        modelId: serial("modelId").primaryKey(),
+        providerId: integer("providerId")
+            .notNull()
+            .references(() => aiProviders.providerId, { onDelete: "cascade" }),
+        modelKey: varchar("modelKey").notNull(),
+        name: varchar("name").notNull(),
+        listType: varchar("listType")
+            .$type<"allow" | "block">()
+            .notNull()
+            .default("allow"),
+        enabled: boolean("enabled").notNull().default(true),
+        createdAt: bigint("createdAt", { mode: "number" }).notNull(),
+        updatedAt: bigint("updatedAt", { mode: "number" }).notNull()
+    },
+    (t) => [unique("ai_model_provider_key_uniq").on(t.providerId, t.modelKey)]
+);
+
+export const aiBudgets = pgTable(
+    "aiBudgets",
+    {
+        budgetId: serial("budgetId").primaryKey(),
+        orgId: varchar("orgId")
+            .notNull()
+            .references(() => orgs.orgId, { onDelete: "cascade" }),
+        providerId: integer("providerId").references(
+            () => aiProviders.providerId,
+            { onDelete: "cascade" }
+        ),
+        modelId: integer("modelId").references(() => aiModels.modelId, {
+            onDelete: "cascade"
+        }),
+        resourceId: integer("resourceId").references(
+            () => resources.resourceId,
+            { onDelete: "cascade" }
+        ),
+        siteResourceId: integer("siteResourceId").references(
+            () => siteResources.siteResourceId,
+            { onDelete: "cascade" }
+        ),
+        roleId: integer("roleId").references(() => roles.roleId, {
+            onDelete: "cascade"
+        }),
+        amount: real("amount").notNull(),
+        unit: varchar("unit").$type<"usd" | "tokens">().notNull(),
+        period: varchar("period")
+            .$type<
+                | "monthly"
+                | "yearly"
+                | "lifetime"
+                | "daily"
+                | "hourly"
+                | "weekly"
+            >()
+            .notNull()
+            .default("monthly"),
+        enforcement: varchar("enforcement")
+            .$type<"hard" | "soft">()
+            .notNull()
+            .default("hard"),
+        enabled: boolean("enabled").notNull().default(true),
+        createdAt: bigint("createdAt", { mode: "number" }).notNull(),
+        updatedAt: bigint("updatedAt", { mode: "number" }).notNull()
+    },
+    (t) => [
+        unique("ai_budget_provider_uniq").on(t.providerId, t.unit, t.period),
+        unique("ai_budget_model_uniq").on(t.modelId, t.unit, t.period),
+        unique("ai_budget_resource_uniq").on(t.resourceId, t.unit, t.period),
+        unique("ai_budget_site_resource_uniq").on(
+            t.siteResourceId,
+            t.unit,
+            t.period
+        ),
+        unique("ai_budget_role_uniq").on(t.roleId, t.unit, t.period)
+    ]
+);
+
 export type Org = InferSelectModel<typeof orgs>;
 export type User = InferSelectModel<typeof users>;
 export type Site = InferSelectModel<typeof sites>;
@@ -1624,3 +1842,12 @@ export type ResourcePolicy = InferSelectModel<typeof resourcePolicies>;
 export type RolePolicy = InferSelectModel<typeof rolePolicies>;
 export type UserPolicy = InferSelectModel<typeof userPolicies>;
 export type ResourcePolicyRule = InferSelectModel<typeof resourcePolicyRules>;
+export type AiProvider = InferSelectModel<typeof aiProviders>;
+export type AiModel = InferSelectModel<typeof aiModels>;
+export type AiBudget = InferSelectModel<typeof aiBudgets>;
+export type ResourceAiProvider = InferSelectModel<typeof resourceAiProviders>;
+export type SiteResourceAiProvider = InferSelectModel<
+    typeof siteResourceAiProviders
+>;
+export type ResourceAiModel = InferSelectModel<typeof resourceAiModels>;
+export type SiteResourceAiModel = InferSelectModel<typeof siteResourceAiModels>;

server/db/sqlite/driver.ts

@@ -5,6 +5,7 @@ import path from "path";
 import fs from "fs";
 import { APP_PATH } from "@server/lib/consts";
 import { existsSync, mkdirSync } from "fs";
+import logger from "@server/logger";
 
 export const location = path.join(APP_PATH, "db", "db.sqlite");
 export const exists = checkFileExists(location);
@@ -12,7 +13,11 @@ export const exists = checkFileExists(location);
 bootstrapVolume();
 
 function createDb() {
-    const sqlite = new Database(location);
+    const verbose =
+        process.env.QUERY_LOGGING == "true"
+            ? (message: unknown) => logger.debug(String(message))
+            : undefined;
+    const sqlite = new Database(location, { verbose });
 
     if (process.env.ENABLE_SQLITE_WAL_MODE == "true") {
         // Enable WAL mode — allows concurrent readers + single writer, preventing

server/db/sqlite/schema/schema.ts

@@ -1,9 +1,11 @@
 import { randomUUID } from "crypto";
-import { InferSelectModel } from "drizzle-orm";
+import { InferSelectModel, sql } from "drizzle-orm";
 import {
+    check,
     index,
     integer,
     primaryKey,
+    real,
     sqliteTable,
     text,
     unique
@@ -107,7 +109,7 @@ export const sites = sqliteTable("sites", {
     }),
     name: text("name").notNull(),
     pubKey: text("pubKey"),
-    subnet: text("subnet"),
+    exitNodeSubnet: text("exitNodeSubnet"),
     megabytesIn: integer("bytesIn").default(0),
     megabytesOut: integer("bytesOut").default(0),
     lastBandwidthUpdate: text("lastBandwidthUpdate"),
@@ -203,7 +205,10 @@ export const resources = sqliteTable("resources", {
     postAuthPath: text("postAuthPath"),
     health: text("health").default("unknown"), // "healthy", "unhealthy", "unknown"
     wildcard: integer("wildcard", { mode: "boolean" }).notNull().default(false),
-    mode: text("mode").default("http").notNull(), // rdp, ssh, http, vnc
+    mode: text("mode")
+        .default("http")
+        .$type<"rdp" | "ssh" | "http" | "vnc" | "inference" | "tcp" | "udp">()
+        .notNull(), // rdp, ssh, http, vnc, inference
     pamMode: text("pamMode")
         .$type<"passthrough" | "push">()
         .default("passthrough"),
@@ -214,6 +219,41 @@ export const resources = sqliteTable("resources", {
     status: text("status").$type<"pending" | "approved">().default("approved")
 });
 
+export const resourceAiProviders = sqliteTable(
+    "resourceAiProviders",
+    {
+        resourceId: integer("resourceId")
+            .notNull()
+            .references(() => resources.resourceId, { onDelete: "cascade" }),
+        providerId: integer("providerId")
+            .notNull()
+            .references(() => aiProviders.providerId, { onDelete: "cascade" }),
+        accessMode: text("accessMode")
+            .$type<"inherit" | "select">()
+            .notNull()
+            .default("inherit"),
+        enabled: integer("enabled", { mode: "boolean" }).notNull().default(true)
+    },
+    (t) => [primaryKey({ columns: [t.resourceId, t.providerId] })]
+);
+
+export const resourceAiModels = sqliteTable(
+    "resourceAiModels",
+    {
+        resourceId: integer("resourceId")
+            .notNull()
+            .references(() => resources.resourceId, { onDelete: "cascade" }),
+        modelId: integer("modelId")
+            .notNull()
+            .references(() => aiModels.modelId, { onDelete: "cascade" }),
+        listType: text("listType")
+            .$type<"allow" | "block">()
+            .notNull()
+            .default("allow")
+    },
+    (t) => [primaryKey({ columns: [t.resourceId, t.modelId] })]
+);
+
 export const labels = sqliteTable("labels", {
     labelId: integer("labelId").primaryKey({ autoIncrement: true }),
     name: text("name").notNull(),
@@ -322,11 +362,12 @@ export const clientLabels = sqliteTable(
 
 export const targets = sqliteTable("targets", {
     targetId: integer("targetId").primaryKey({ autoIncrement: true }),
-    resourceId: integer("resourceId")
-        .references(() => resources.resourceId, {
-            onDelete: "cascade"
-        })
-        .notNull(),
+    resourceId: integer("resourceId").references(() => resources.resourceId, {
+        onDelete: "cascade"
+    }),
+    providerId: integer("providerId").references(() => aiProviders.providerId, {
+        onDelete: "cascade"
+    }),
     siteId: integer("siteId")
         .references(() => sites.siteId, {
             onDelete: "cascade"
@@ -422,10 +463,17 @@ export const siteResources = sqliteTable("siteResources", {
         () => networks.networkId,
         { onDelete: "restrict" }
     ),
+    requiresExitNodeConnection: integer("requiresExitNodeConnection", {
+        mode: "boolean"
+    })
+        .notNull()
+        .default(false),
     niceId: text("niceId").notNull(),
     name: text("name").notNull(),
     ssl: integer("ssl", { mode: "boolean" }).notNull().default(false),
-    mode: text("mode").$type<"host" | "cidr" | "http" | "ssh">().notNull(), // "host" | "cidr" | "http"
+    mode: text("mode")
+        .$type<"host" | "cidr" | "http" | "ssh" | "inference">()
+        .notNull(), // "host" | "cidr" | "http"
     scheme: text("scheme").$type<"http" | "https">(), // only for when we are doing https or http mode
     proxyPort: integer("proxyPort"), // only for port mode
     destinationPort: integer("destinationPort"), // only for port mode
@@ -453,6 +501,45 @@ export const siteResources = sqliteTable("siteResources", {
     status: text("status").$type<"pending" | "approved">().default("approved")
 });
 
+export const siteResourceAiProviders = sqliteTable(
+    "siteResourceAiProviders",
+    {
+        siteResourceId: integer("siteResourceId")
+            .notNull()
+            .references(() => siteResources.siteResourceId, {
+                onDelete: "cascade"
+            }),
+        providerId: integer("providerId")
+            .notNull()
+            .references(() => aiProviders.providerId, { onDelete: "cascade" }),
+        accessMode: text("accessMode")
+            .$type<"inherit" | "select">()
+            .notNull()
+            .default("inherit"),
+        enabled: integer("enabled", { mode: "boolean" }).notNull().default(true)
+    },
+    (t) => [primaryKey({ columns: [t.siteResourceId, t.providerId] })]
+);
+
+export const siteResourceAiModels = sqliteTable(
+    "siteResourceAiModels",
+    {
+        siteResourceId: integer("siteResourceId")
+            .notNull()
+            .references(() => siteResources.siteResourceId, {
+                onDelete: "cascade"
+            }),
+        modelId: integer("modelId")
+            .notNull()
+            .references(() => aiModels.modelId, { onDelete: "cascade" }),
+        listType: text("listType")
+            .$type<"allow" | "block">()
+            .notNull()
+            .default("allow")
+    },
+    (t) => [primaryKey({ columns: [t.siteResourceId, t.modelId] })]
+);
+
 export const networks = sqliteTable("networks", {
     networkId: integer("networkId").primaryKey({ autoIncrement: true }),
     niceId: text("niceId"),
@@ -599,6 +686,7 @@ export const clients = sqliteTable("clients", {
     pubKey: text("pubKey"),
     olmId: text("olmId"), // to lock it to a specific olm optionally
     subnet: text("subnet").notNull(),
+    exitNodeSubnet: text("exitNodeSubnet"), // this is the subnet when connecting to an exit node
     megabytesIn: integer("bytesIn"),
     megabytesOut: integer("bytesOut"),
     lastBandwidthUpdate: text("lastBandwidthUpdate"),
@@ -1526,6 +1614,136 @@ export const statusHistory = sqliteTable(
     ]
 );
 
+export const aiProviders = sqliteTable("aiProviders", {
+    providerId: integer("providerId").primaryKey({ autoIncrement: true }),
+    orgId: text("orgId")
+        .notNull()
+        .references(() => orgs.orgId, { onDelete: "cascade" }),
+    name: text("name").notNull(),
+    type: text("type")
+        .$type<
+            | "openai"
+            | "anthropic"
+            | "googleGemini"
+            | "vertexAi"
+            | "bedrock"
+            | "microsoftFoundry"
+            | "openRouter"
+            | "vercelAiGateway"
+            | "custom"
+        >()
+        .notNull(),
+    upstreamUrl: text("upstreamUrl"),
+    apiKey: text("apiKey"),
+    apiKeyLastChars: text("apiKeyLastChars"),
+    authType: text("authType")
+        .$type<
+            | "bearer"
+            | "x-api-key"
+            | "x-goog-api-key"
+            | "hec"
+            | "cf-aig-authorization"
+            | "none"
+            | "passthrough"
+        >()
+        .notNull(),
+    routingMode: text("routingMode")
+        .$type<"url" | "target">()
+        .notNull()
+        .default("url"),
+    capabilities: text("capabilities").notNull().default("[]"),
+    headers: text("headers"), // JSON array of { name, value }
+    skipTlsVerification: integer("skipTlsVerification", { mode: "boolean" })
+        .notNull()
+        .default(false),
+    enabled: integer("enabled", { mode: "boolean" }).notNull().default(true),
+    createdAt: integer("createdAt").notNull(),
+    updatedAt: integer("updatedAt").notNull()
+});
+
+export const aiModels = sqliteTable(
+    "aiModels",
+    {
+        modelId: integer("modelId").primaryKey({ autoIncrement: true }),
+        providerId: integer("providerId")
+            .notNull()
+            .references(() => aiProviders.providerId, { onDelete: "cascade" }),
+        modelKey: text("modelKey").notNull(),
+        name: text("name").notNull(),
+        listType: text("listType")
+            .$type<"allow" | "block">()
+            .notNull()
+            .default("allow"),
+        enabled: integer("enabled", { mode: "boolean" })
+            .notNull()
+            .default(true),
+        createdAt: integer("createdAt").notNull(),
+        updatedAt: integer("updatedAt").notNull()
+    },
+    (t) => [unique("ai_model_provider_key_uniq").on(t.providerId, t.modelKey)]
+);
+
+export const aiBudgets = sqliteTable(
+    "aiBudgets",
+    {
+        budgetId: integer("budgetId").primaryKey({ autoIncrement: true }),
+        orgId: text("orgId")
+            .notNull()
+            .references(() => orgs.orgId, { onDelete: "cascade" }),
+        providerId: integer("providerId").references(
+            () => aiProviders.providerId,
+            { onDelete: "cascade" }
+        ),
+        modelId: integer("modelId").references(() => aiModels.modelId, {
+            onDelete: "cascade"
+        }),
+        resourceId: integer("resourceId").references(
+            () => resources.resourceId,
+            { onDelete: "cascade" }
+        ),
+        siteResourceId: integer("siteResourceId").references(
+            () => siteResources.siteResourceId,
+            { onDelete: "cascade" }
+        ),
+        roleId: integer("roleId").references(() => roles.roleId, {
+            onDelete: "cascade"
+        }),
+        amount: real("amount").notNull(),
+        unit: text("unit").$type<"usd" | "tokens">().notNull(),
+        period: text("period")
+            .$type<
+                | "monthly"
+                | "yearly"
+                | "lifetime"
+                | "daily"
+                | "hourly"
+                | "weekly"
+            >()
+            .notNull()
+            .default("monthly"),
+        enforcement: text("enforcement")
+            .$type<"hard" | "soft">()
+            .notNull()
+            .default("hard"),
+        enabled: integer("enabled", { mode: "boolean" })
+            .notNull()
+            .default(true),
+        createdAt: integer("createdAt").notNull(),
+        updatedAt: integer("updatedAt").notNull()
+    },
+    (t) => [
+        unique("ai_budget_provider_uniq").on(t.providerId, t.unit, t.period),
+        unique("ai_budget_model_uniq").on(t.modelId, t.unit, t.period),
+        unique("ai_budget_resource_uniq").on(t.resourceId, t.unit, t.period),
+        unique("ai_budget_site_resource_uniq").on(
+            t.siteResourceId,
+            t.unit,
+            t.period
+        ),
+        unique("ai_budget_role_uniq").on(t.roleId, t.unit, t.period)
+    ]
+);
+
 export type Org = InferSelectModel<typeof orgs>;
 export type User = InferSelectModel<typeof users>;
 export type Site = InferSelectModel<typeof sites>;
@@ -1608,3 +1826,12 @@ export type ResourcePolicyHeaderAuth = InferSelectModel<
 >;
 export type RolePolicy = InferSelectModel<typeof rolePolicies>;
 export type UserPolicy = InferSelectModel<typeof userPolicies>;
+export type AiProvider = InferSelectModel<typeof aiProviders>;
+export type AiModel = InferSelectModel<typeof aiModels>;
+export type AiBudget = InferSelectModel<typeof aiBudgets>;
+export type ResourceAiProvider = InferSelectModel<typeof resourceAiProviders>;
+export type SiteResourceAiProvider = InferSelectModel<
+    typeof siteResourceAiProviders
+>;
+export type ResourceAiModel = InferSelectModel<typeof resourceAiModels>;
+export type SiteResourceAiModel = InferSelectModel<typeof siteResourceAiModels>;

server/index.ts

@@ -5,10 +5,14 @@ import { runSetupFunctions } from "./setup";
 import { createApiServer } from "./apiServer";
 import { createNextServer } from "./nextServer";
 import { createInternalServer } from "./internalServer";
+import { createAiGatewayServer } from "./aiGatewayServer";
 import { createIntegrationApiServer } from "./integrationApiServer";
 import {
     ApiKey,
     ApiKeyOrg,
+    AiBudget,
+    AiModel,
+    AiProvider,
     RemoteExitNode,
     Session,
     SiteResource,
@@ -25,6 +29,7 @@ import { initLogCleanupInterval } from "@server/lib/cleanupLogs";
 import { initAcmeCertSync } from "#dynamic/lib/acmeCertSync";
 import { fetchServerIp } from "@server/lib/serverIpService";
 import { startRebuildQueueProcessor } from "@server/lib/rebuildClientAssociations";
+import { initAiModelCatalog } from "@server/lib/aiModelCatalog";
 
 async function startServers() {
     await setHostMeta();
@@ -43,10 +48,12 @@ async function startServers() {
     initLogCleanupInterval();
     initAcmeCertSync();
     startRebuildQueueProcessor();
+    await initAiModelCatalog();
 
     // Start all servers
     const apiServer = createApiServer();
     const internalServer = createInternalServer();
+    const aiGatewayServer = createAiGatewayServer();
 
     const nextServer = await createNextServer();
     if (config.getRawConfig().traefik.file_mode) {
@@ -65,6 +72,7 @@ async function startServers() {
         apiServer,
         nextServer,
         internalServer,
+        aiGatewayServer,
         integrationServer
     };
 }
@@ -83,6 +91,9 @@ declare global {
             userOrgIds?: string[];
             remoteExitNode?: RemoteExitNode;
             siteResource?: SiteResource;
+            aiProvider?: AiProvider;
+            aiModel?: AiModel;
+            aiBudget?: AiBudget;
             orgPolicyAllowed?: boolean;
         }
     }

server/lib/aiCapabilities.ts

@@ -0,0 +1,283 @@
+import type { Request } from "express";
+
+export const AI_CAPABILITIES = [
+    "openai_chat",
+    "openai_responses",
+    "anthropic_messages",
+    "gemini_generate_content",
+    "bedrock_model_invoke",
+    "google_generate_content",
+    "google_raw_predict",
+    "bedrock_converse"
+] as const;
+
+export type AiCapability = (typeof AI_CAPABILITIES)[number];
+
+export type AiCapabilityRoute = {
+    method: "POST";
+    path: string;
+};
+
+export type AiCapabilityDefinition = {
+    id: AiCapability;
+    routes: AiCapabilityRoute[];
+    extractModel: (req: Request) => string | undefined;
+    resolveUpstreamUrl: (
+        baseUrl: string,
+        req: Request,
+        model: string
+    ) => string;
+    isStreaming: (req: Request, contentType: string) => boolean;
+};
+
+function bodyModel(req: Request): string | undefined {
+    return typeof req.body?.model === "string" ? req.body.model : undefined;
+}
+
+function paramModel(req: Request): string | undefined {
+    const model = req.params?.model;
+    return typeof model === "string" && model.length > 0 ? model : undefined;
+}
+
+export function joinUpstreamUrl(baseUrl: string, path: string): string {
+    const base = baseUrl.replace(/\/+$/, "");
+    let suffix = path.startsWith("/") ? path : `/${path}`;
+
+    let basePathname = "/";
+    try {
+        basePathname = new URL(base).pathname.replace(/\/+$/, "") || "/";
+    } catch {
+        // Fall through with "/" non-absolute bases are not expected in
+        // production, but keep joining usable for malformed input.
+    }
+
+    if (basePathname !== "/") {
+        const baseSegs = basePathname.split("/").filter(Boolean);
+        const pathSegs = suffix.split("/").filter(Boolean);
+        const max = Math.min(baseSegs.length, pathSegs.length);
+        let overlap = 0;
+        for (let n = max; n >= 1; n--) {
+            const baseSuffix = baseSegs.slice(-n);
+            const pathPrefix = pathSegs.slice(0, n);
+            if (baseSuffix.every((seg, i) => seg === pathPrefix[i])) {
+                overlap = n;
+                break;
+            }
+        }
+        if (overlap > 0) {
+            const remaining = pathSegs.slice(overlap);
+            suffix = remaining.length > 0 ? `/${remaining.join("/")}` : "/";
+        }
+    }
+
+    if (suffix === "/") {
+        return base;
+    }
+
+    return `${base}${suffix}`;
+}
+
+function pathFromRequest(req: Request): string {
+    const raw = req.originalUrl || req.url || req.path;
+    return raw.startsWith("/") ? raw : `/${raw}`;
+}
+
+function bodyRequestsStream(req: Request): boolean {
+    return req.body?.stream === true;
+}
+
+function contentTypeIsSse(contentType: string): boolean {
+    return contentType.includes("text/event-stream");
+}
+
+function contentTypeIsAmazonEventStream(contentType: string): boolean {
+    return contentType.includes("application/vnd.amazon.eventstream");
+}
+
+function pathIncludes(req: Request, fragment: string): boolean {
+    return pathFromRequest(req).includes(fragment);
+}
+
+function isBodyOrSseStreaming(req: Request, contentType: string): boolean {
+    return bodyRequestsStream(req) || contentTypeIsSse(contentType);
+}
+
+function isGeminiStyleStreaming(req: Request, contentType: string): boolean {
+    return (
+        pathIncludes(req, "streamGenerateContent") ||
+        pathIncludes(req, "alt=sse") ||
+        contentTypeIsSse(contentType)
+    );
+}
+
+export const AI_CAPABILITY_DEFS: Record<AiCapability, AiCapabilityDefinition> =
+    {
+        openai_chat: {
+            id: "openai_chat",
+            routes: [
+                { method: "POST", path: "/v1/chat/completions" },
+                { method: "POST", path: "/chat/completions" }
+            ],
+            extractModel: bodyModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: isBodyOrSseStreaming
+        },
+        openai_responses: {
+            id: "openai_responses",
+            routes: [{ method: "POST", path: "/v1/responses" }],
+            extractModel: bodyModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: isBodyOrSseStreaming
+        },
+        anthropic_messages: {
+            id: "anthropic_messages",
+            routes: [{ method: "POST", path: "/v1/messages" }],
+            extractModel: bodyModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: isBodyOrSseStreaming
+        },
+        gemini_generate_content: {
+            id: "gemini_generate_content",
+            routes: [
+                {
+                    method: "POST",
+                    path: "/v1beta/models/:model\\:generateContent"
+                },
+                {
+                    method: "POST",
+                    path: "/v1beta/models/:model\\:streamGenerateContent"
+                }
+            ],
+            extractModel: paramModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: isGeminiStyleStreaming
+        },
+        google_generate_content: {
+            id: "google_generate_content",
+            routes: [
+                {
+                    method: "POST",
+                    // Vertex publisher model generateContent
+                    path: "/v1/projects/:project/locations/:location/publishers/:publisher/models/:model\\:generateContent"
+                },
+                {
+                    method: "POST",
+                    path: "/v1/projects/:project/locations/:location/publishers/:publisher/models/:model\\:streamGenerateContent"
+                }
+            ],
+            extractModel: paramModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: isGeminiStyleStreaming
+        },
+        google_raw_predict: {
+            id: "google_raw_predict",
+            routes: [
+                {
+                    method: "POST",
+                    path: "/v1/projects/:project/locations/:location/publishers/:publisher/models/:model\\:rawPredict"
+                },
+                {
+                    method: "POST",
+                    path: "/v1/projects/:project/locations/:location/publishers/:publisher/models/:model\\:streamRawPredict"
+                }
+            ],
+            extractModel: paramModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: (req, contentType) =>
+                pathIncludes(req, "streamRawPredict") ||
+                pathIncludes(req, "alt=sse") ||
+                contentTypeIsSse(contentType)
+        },
+        bedrock_model_invoke: {
+            id: "bedrock_model_invoke",
+            routes: [
+                { method: "POST", path: "/model/:model/invoke" },
+                {
+                    method: "POST",
+                    path: "/model/:model/invoke-with-response-stream"
+                }
+            ],
+            extractModel: paramModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: (req, contentType) =>
+                pathIncludes(req, "invoke-with-response-stream") ||
+                contentTypeIsAmazonEventStream(contentType) ||
+                contentTypeIsSse(contentType)
+        },
+        bedrock_converse: {
+            id: "bedrock_converse",
+            routes: [
+                { method: "POST", path: "/model/:model/converse" },
+                { method: "POST", path: "/model/:model/converse-stream" }
+            ],
+            extractModel: paramModel,
+            resolveUpstreamUrl: (base, req) =>
+                joinUpstreamUrl(base, pathFromRequest(req)),
+            isStreaming: (req, contentType) =>
+                pathIncludes(req, "converse-stream") ||
+                contentTypeIsAmazonEventStream(contentType) ||
+                contentTypeIsSse(contentType)
+        }
+    };
+
+export function isAiCapability(value: unknown): value is AiCapability {
+    return (
+        typeof value === "string" &&
+        (AI_CAPABILITIES as readonly string[]).includes(value)
+    );
+}
+
+export function parseCapabilities(raw: unknown): AiCapability[] {
+    if (raw == null) {
+        return [];
+    }
+
+    let parsed: unknown = raw;
+    if (typeof raw === "string") {
+        const trimmed = raw.trim();
+        if (!trimmed) {
+            return [];
+        }
+        try {
+            parsed = JSON.parse(trimmed);
+        } catch {
+            return [];
+        }
+    }
+
+    if (!Array.isArray(parsed)) {
+        return [];
+    }
+
+    const out: AiCapability[] = [];
+    const seen = new Set<AiCapability>();
+    for (const item of parsed) {
+        if (isAiCapability(item) && !seen.has(item)) {
+            seen.add(item);
+            out.push(item);
+        }
+    }
+    return out;
+}
+
+export function serializeCapabilities(capabilities: AiCapability[]): string {
+    return JSON.stringify(capabilities);
+}
+
+export function providerHasCapability(
+    capabilities: AiCapability[] | string | null | undefined,
+    capability: AiCapability
+): boolean {
+    const list =
+        typeof capabilities === "string" || capabilities == null
+            ? parseCapabilities(capabilities)
+            : capabilities;
+    return list.includes(capability);
+}

server/lib/aiGatewayUpstreamFetch.ts

@@ -0,0 +1,82 @@
+import http from "node:http";
+import https from "node:https";
+import { Readable } from "node:stream";
+
+type UpstreamFetchInit = {
+    method: string;
+    headers: Record<string, string>;
+    body?: string;
+    skipTlsVerification?: boolean;
+    signal?: AbortSignal;
+};
+
+const insecureHttpsAgent = new https.Agent({
+    rejectUnauthorized: false,
+    keepAlive: true
+});
+
+export function aiGatewayUpstreamFetch(
+    url: string,
+    init: UpstreamFetchInit
+): Promise<Response> {
+    const parsed = new URL(url);
+    const isHttps = parsed.protocol === "https:";
+    const lib = isHttps ? https : http;
+    const agent =
+        isHttps && init.skipTlsVerification ? insecureHttpsAgent : undefined;
+
+    return new Promise((resolve, reject) => {
+        if (init.signal?.aborted) {
+            reject(init.signal.reason ?? new Error("Request aborted"));
+            return;
+        }
+
+        const req = lib.request(
+            url,
+            {
+                method: init.method,
+                headers: init.headers,
+                agent
+            },
+            (res) => {
+                const headers = new Headers();
+                for (const [key, value] of Object.entries(res.headers)) {
+                    if (value === undefined) {
+                        continue;
+                    }
+                    if (Array.isArray(value)) {
+                        for (const entry of value) {
+                            headers.append(key, entry);
+                        }
+                    } else {
+                        headers.set(key, value);
+                    }
+                }
+
+                const body = Readable.toWeb(res) as ReadableStream<Uint8Array>;
+                resolve(
+                    new Response(body, {
+                        status: res.statusCode ?? 502,
+                        statusText: res.statusMessage,
+                        headers
+                    })
+                );
+            }
+        );
+
+        req.on("error", reject);
+
+        if (init.signal) {
+            const onAbort = () => req.destroy(init.signal!.reason);
+            init.signal.addEventListener("abort", onAbort, { once: true });
+            req.on("close", () =>
+                init.signal!.removeEventListener("abort", onAbort)
+            );
+        }
+
+        if (init.body !== undefined) {
+            req.write(init.body);
+        }
+        req.end();
+    });
+}

server/lib/aiInferenceResource.ts

@@ -0,0 +1,734 @@
+import { and, eq, inArray } from "drizzle-orm";
+import {
+    aiModels,
+    aiProviders,
+    db,
+    resourceAiModels,
+    resourceAiProviders,
+    siteResourceAiModels,
+    siteResourceAiProviders,
+    type Transaction
+} from "@server/db";
+import { z } from "zod";
+import { modelKeysConflict } from "@server/lib/aiModelKeyMatch";
+
+type DbOrTrx = Transaction | typeof db;
+
+export const modelListTypeSchema = z.enum(["allow", "block"]);
+
+export type ModelListType = z.infer<typeof modelListTypeSchema>;
+
+export const accessModeSchema = z.enum(["inherit", "select"]);
+
+export type AccessMode = z.infer<typeof accessModeSchema>;
+
+export const resourceAiProviderAttachmentSchema = z.strictObject({
+    providerId: z.number().int().positive(),
+    accessMode: accessModeSchema.optional().default("inherit"),
+    enabled: z.boolean().optional().default(true)
+});
+
+export type ResourceAiProviderInput = z.infer<
+    typeof resourceAiProviderAttachmentSchema
+>;
+
+export type ResourceAiProviderAttachment = {
+    providerId: number;
+    accessMode: AccessMode;
+    enabled: boolean;
+};
+
+export const resourceAiModelEntrySchema = z.strictObject({
+    modelId: z.number().int().positive(),
+    listType: modelListTypeSchema
+});
+
+export type ResourceAiModelEntry = z.infer<typeof resourceAiModelEntrySchema>;
+
+export type InferenceFieldsError = {
+    error: string;
+};
+
+export function isInferenceFieldsError(
+    value: { error: string } | object
+): value is InferenceFieldsError {
+    return "error" in value;
+}
+
+/**
+ * Resolve which allow/block patterns apply for an attachment.
+ * inherit → provider lists; select → resource-selected lists (replace).
+ */
+export function resolveEffectiveLists(input: {
+    accessMode: AccessMode;
+    providerAllows: string[];
+    providerBlocks: string[];
+    resourceAllows: string[];
+    resourceBlocks: string[];
+}): { allows: string[]; blocks: string[] } {
+    if (input.accessMode === "select") {
+        return {
+            allows: input.resourceAllows,
+            blocks: input.resourceBlocks
+        };
+    }
+    return {
+        allows: input.providerAllows,
+        blocks: input.providerBlocks
+    };
+}
+
+function normalizeAttachments(
+    inputs: ResourceAiProviderInput[]
+): ResourceAiProviderAttachment[] {
+    const byProviderId = new Map<
+        number,
+        { accessMode: AccessMode; enabled: boolean }
+    >();
+    for (const input of inputs) {
+        byProviderId.set(input.providerId, {
+            accessMode: input.accessMode ?? "inherit",
+            enabled: input.enabled ?? true
+        });
+    }
+    return [...byProviderId.entries()].map(
+        ([providerId, { accessMode, enabled }]) => ({
+            providerId,
+            accessMode,
+            enabled
+        })
+    );
+}
+
+type EffectiveAllowRow = {
+    providerId: number;
+    modelKey: string;
+};
+
+/**
+ * Ensure effective allow modelKeys do not conflict across attached providers.
+ * inherit uses provider allows; select uses resource-selected allows (or the
+ * optional override map). Block patterns are ignored for overlap checks.
+ */
+export async function assertNoOverlappingModelKeys(
+    attachments: ResourceAiProviderAttachment[],
+    options: {
+        trx?: DbOrTrx;
+        resourceId?: number;
+        siteResourceId?: number;
+        selectedAllowsByProvider?: Map<number, string[]>;
+    } = {}
+): Promise<InferenceFieldsError | null> {
+    const trx = options.trx ?? db;
+
+    const activeAttachments = attachments.filter((a) => a.enabled);
+
+    if (activeAttachments.length < 2) {
+        return null;
+    }
+
+    const inheritProviderIds = activeAttachments
+        .filter((a) => a.accessMode === "inherit")
+        .map((a) => a.providerId);
+    const selectProviderIds = activeAttachments
+        .filter((a) => a.accessMode === "select")
+        .map((a) => a.providerId);
+
+    const effectiveAllows: EffectiveAllowRow[] = [];
+
+    if (inheritProviderIds.length > 0) {
+        const providerAllows = await trx
+            .select({
+                providerId: aiModels.providerId,
+                modelKey: aiModels.modelKey
+            })
+            .from(aiModels)
+            .where(
+                and(
+                    inArray(aiModels.providerId, inheritProviderIds),
+                    eq(aiModels.enabled, true),
+                    eq(aiModels.listType, "allow")
+                )
+            );
+        effectiveAllows.push(...providerAllows);
+    }
+
+    if (selectProviderIds.length > 0) {
+        if (options.selectedAllowsByProvider) {
+            for (const providerId of selectProviderIds) {
+                const keys =
+                    options.selectedAllowsByProvider.get(providerId) ?? [];
+                for (const modelKey of keys) {
+                    effectiveAllows.push({ providerId, modelKey });
+                }
+            }
+        } else if (options.resourceId !== undefined) {
+            const rows = await trx
+                .select({
+                    providerId: aiModels.providerId,
+                    modelKey: aiModels.modelKey
+                })
+                .from(resourceAiModels)
+                .innerJoin(
+                    aiModels,
+                    eq(resourceAiModels.modelId, aiModels.modelId)
+                )
+                .where(
+                    and(
+                        eq(resourceAiModels.resourceId, options.resourceId),
+                        inArray(aiModels.providerId, selectProviderIds),
+                        eq(resourceAiModels.listType, "allow"),
+                        eq(aiModels.enabled, true)
+                    )
+                );
+            effectiveAllows.push(...rows);
+        } else if (options.siteResourceId !== undefined) {
+            const rows = await trx
+                .select({
+                    providerId: aiModels.providerId,
+                    modelKey: aiModels.modelKey
+                })
+                .from(siteResourceAiModels)
+                .innerJoin(
+                    aiModels,
+                    eq(siteResourceAiModels.modelId, aiModels.modelId)
+                )
+                .where(
+                    and(
+                        eq(
+                            siteResourceAiModels.siteResourceId,
+                            options.siteResourceId
+                        ),
+                        inArray(aiModels.providerId, selectProviderIds),
+                        eq(siteResourceAiModels.listType, "allow"),
+                        eq(aiModels.enabled, true)
+                    )
+                );
+            effectiveAllows.push(...rows);
+        }
+    }
+
+    const conflictPairs: string[] = [];
+    for (let i = 0; i < effectiveAllows.length; i++) {
+        for (let j = i + 1; j < effectiveAllows.length; j++) {
+            const left = effectiveAllows[i];
+            const right = effectiveAllows[j];
+            if (left.providerId === right.providerId) {
+                continue;
+            }
+            if (!modelKeysConflict(left.modelKey, right.modelKey)) {
+                continue;
+            }
+            const pair = [left.modelKey, right.modelKey].sort().join(" vs ");
+            if (!conflictPairs.includes(pair)) {
+                conflictPairs.push(pair);
+            }
+        }
+    }
+
+    if (conflictPairs.length === 0) {
+        return null;
+    }
+
+    conflictPairs.sort();
+    return {
+        error: `Model keys must be unique across providers on a resource. Overlapping keys: ${conflictPairs.join(", ")}`
+    };
+}
+
+/**
+ * Validate provider attachments for an org.
+ */
+export async function resolveProviderAttachments(input: {
+    orgId: string;
+    attachments: ResourceAiProviderInput[];
+    requireAtLeastOne: boolean;
+    resourceId?: number;
+    siteResourceId?: number;
+}): Promise<ResourceAiProviderAttachment[] | InferenceFieldsError> {
+    const attachments = normalizeAttachments(input.attachments);
+
+    if (input.requireAtLeastOne && attachments.length === 0) {
+        return {
+            error: "At least one AI provider is required for inference-mode resources"
+        };
+    }
+
+    if (attachments.length === 0) {
+        return [];
+    }
+
+    const providerIds = attachments.map((a) => a.providerId);
+    const providers = await db
+        .select({
+            providerId: aiProviders.providerId,
+            orgId: aiProviders.orgId,
+            enabled: aiProviders.enabled
+        })
+        .from(aiProviders)
+        .where(
+            and(
+                inArray(aiProviders.providerId, providerIds),
+                eq(aiProviders.orgId, input.orgId)
+            )
+        );
+
+    if (providers.length !== providerIds.length) {
+        return {
+            error: "One or more AI providers were not found in this organization"
+        };
+    }
+
+    const disabled = providers.find((p) => !p.enabled);
+    if (disabled) {
+        return {
+            error: `AI provider with ID ${disabled.providerId} is disabled`
+        };
+    }
+
+    const overlapError = await assertNoOverlappingModelKeys(attachments, {
+        resourceId: input.resourceId,
+        siteResourceId: input.siteResourceId
+    });
+    if (overlapError) {
+        return overlapError;
+    }
+
+    return attachments;
+}
+
+export async function assertInferenceModeAllowsProviderFields(input: {
+    mode: string;
+    hasProviderAttachments: boolean;
+}): Promise<InferenceFieldsError | null> {
+    if (input.mode === "inference") {
+        return null;
+    }
+    if (input.hasProviderAttachments) {
+        return {
+            error: "AI providers can only be attached to inference-mode resources"
+        };
+    }
+    return null;
+}
+
+/**
+ * Attach providers to a resource. Inherit attachments use the provider lists
+ * as-is (resource model rows for those providers are pruned). Select
+ * attachments keep resource-selected allow/block subsets.
+ */
+export async function setPublicResourceAiProviders(
+    resourceId: number,
+    attachments: ResourceAiProviderAttachment[],
+    trx: DbOrTrx = db
+): Promise<void> {
+    await trx
+        .delete(resourceAiProviders)
+        .where(eq(resourceAiProviders.resourceId, resourceId));
+
+    if (attachments.length > 0) {
+        await trx.insert(resourceAiProviders).values(
+            attachments.map((a) => ({
+                resourceId,
+                providerId: a.providerId,
+                accessMode: a.accessMode,
+                enabled: a.enabled
+            }))
+        );
+    }
+
+    await prunePublicResourceModelsToSelectProviders(
+        resourceId,
+        attachments,
+        trx
+    );
+}
+
+export async function setSiteResourceAiProviders(
+    siteResourceId: number,
+    attachments: ResourceAiProviderAttachment[],
+    trx: DbOrTrx = db
+): Promise<void> {
+    await trx
+        .delete(siteResourceAiProviders)
+        .where(eq(siteResourceAiProviders.siteResourceId, siteResourceId));
+
+    if (attachments.length > 0) {
+        await trx.insert(siteResourceAiProviders).values(
+            attachments.map((a) => ({
+                siteResourceId,
+                providerId: a.providerId,
+                accessMode: a.accessMode,
+                enabled: a.enabled
+            }))
+        );
+    }
+
+    await pruneSiteResourceModelsToSelectProviders(
+        siteResourceId,
+        attachments,
+        trx
+    );
+}
+
+/**
+ * Keep resource model rows only for providers in select mode.
+ */
+async function prunePublicResourceModelsToSelectProviders(
+    resourceId: number,
+    attachments: ResourceAiProviderAttachment[],
+    trx: DbOrTrx
+): Promise<void> {
+    const selectProviderIds = attachments
+        .filter((a) => a.accessMode === "select")
+        .map((a) => a.providerId);
+
+    if (selectProviderIds.length === 0) {
+        await trx
+            .delete(resourceAiModels)
+            .where(eq(resourceAiModels.resourceId, resourceId));
+        return;
+    }
+
+    const existing = await trx
+        .select({
+            modelId: resourceAiModels.modelId,
+            providerId: aiModels.providerId
+        })
+        .from(resourceAiModels)
+        .innerJoin(aiModels, eq(resourceAiModels.modelId, aiModels.modelId))
+        .where(eq(resourceAiModels.resourceId, resourceId));
+
+    const allowed = new Set(selectProviderIds);
+    const toRemove = existing
+        .filter((row) => !allowed.has(row.providerId))
+        .map((row) => row.modelId);
+
+    if (toRemove.length > 0) {
+        await trx
+            .delete(resourceAiModels)
+            .where(
+                and(
+                    eq(resourceAiModels.resourceId, resourceId),
+                    inArray(resourceAiModels.modelId, toRemove)
+                )
+            );
+    }
+}
+
+async function pruneSiteResourceModelsToSelectProviders(
+    siteResourceId: number,
+    attachments: ResourceAiProviderAttachment[],
+    trx: DbOrTrx
+): Promise<void> {
+    const selectProviderIds = attachments
+        .filter((a) => a.accessMode === "select")
+        .map((a) => a.providerId);
+
+    if (selectProviderIds.length === 0) {
+        await trx
+            .delete(siteResourceAiModels)
+            .where(eq(siteResourceAiModels.siteResourceId, siteResourceId));
+        return;
+    }
+
+    const existing = await trx
+        .select({
+            modelId: siteResourceAiModels.modelId,
+            providerId: aiModels.providerId
+        })
+        .from(siteResourceAiModels)
+        .innerJoin(aiModels, eq(siteResourceAiModels.modelId, aiModels.modelId))
+        .where(eq(siteResourceAiModels.siteResourceId, siteResourceId));
+
+    const allowed = new Set(selectProviderIds);
+    const toRemove = existing
+        .filter((row) => !allowed.has(row.providerId))
+        .map((row) => row.modelId);
+
+    if (toRemove.length > 0) {
+        await trx
+            .delete(siteResourceAiModels)
+            .where(
+                and(
+                    eq(siteResourceAiModels.siteResourceId, siteResourceId),
+                    inArray(siteResourceAiModels.modelId, toRemove)
+                )
+            );
+    }
+}
+
+export async function clearPublicResourceAiConfig(
+    resourceId: number,
+    trx: DbOrTrx = db
+): Promise<void> {
+    await trx
+        .delete(resourceAiModels)
+        .where(eq(resourceAiModels.resourceId, resourceId));
+    await trx
+        .delete(resourceAiProviders)
+        .where(eq(resourceAiProviders.resourceId, resourceId));
+}
+
+export async function clearSiteResourceAiConfig(
+    siteResourceId: number,
+    trx: DbOrTrx = db
+): Promise<void> {
+    await trx
+        .delete(siteResourceAiModels)
+        .where(eq(siteResourceAiModels.siteResourceId, siteResourceId));
+    await trx
+        .delete(siteResourceAiProviders)
+        .where(eq(siteResourceAiProviders.siteResourceId, siteResourceId));
+}
+
+export async function listPublicResourceAiProviders(resourceId: number) {
+    return db
+        .select({
+            providerId: resourceAiProviders.providerId,
+            name: aiProviders.name,
+            type: aiProviders.type,
+            enabled: resourceAiProviders.enabled,
+            providerEnabled: aiProviders.enabled,
+            accessMode: resourceAiProviders.accessMode
+        })
+        .from(resourceAiProviders)
+        .innerJoin(
+            aiProviders,
+            eq(resourceAiProviders.providerId, aiProviders.providerId)
+        )
+        .where(eq(resourceAiProviders.resourceId, resourceId));
+}
+
+export async function listSiteResourceAiProviders(siteResourceId: number) {
+    return db
+        .select({
+            providerId: siteResourceAiProviders.providerId,
+            name: aiProviders.name,
+            type: aiProviders.type,
+            enabled: siteResourceAiProviders.enabled,
+            providerEnabled: aiProviders.enabled,
+            accessMode: siteResourceAiProviders.accessMode
+        })
+        .from(siteResourceAiProviders)
+        .innerJoin(
+            aiProviders,
+            eq(siteResourceAiProviders.providerId, aiProviders.providerId)
+        )
+        .where(eq(siteResourceAiProviders.siteResourceId, siteResourceId));
+}
+
+/**
+ * Model list APIs require an inference resource with at least one select-mode
+ * attached provider.
+ */
+export async function assertPublicModelListApiEligible(resource: {
+    resourceId: number;
+    mode: string;
+}): Promise<string | null> {
+    if (resource.mode !== "inference") {
+        return "AI model lists are only supported on inference-mode resources";
+    }
+
+    const [row] = await db
+        .select({ providerId: resourceAiProviders.providerId })
+        .from(resourceAiProviders)
+        .where(
+            and(
+                eq(resourceAiProviders.resourceId, resource.resourceId),
+                eq(resourceAiProviders.accessMode, "select")
+            )
+        )
+        .limit(1);
+
+    if (!row) {
+        return "Set at least one attached AI provider to select mode before managing model lists";
+    }
+    return null;
+}
+
+export async function assertSiteModelListApiEligible(siteResource: {
+    siteResourceId: number;
+    mode: string;
+}): Promise<string | null> {
+    if (siteResource.mode !== "inference") {
+        return "AI model lists are only supported on inference-mode resources";
+    }
+
+    const [row] = await db
+        .select({ providerId: siteResourceAiProviders.providerId })
+        .from(siteResourceAiProviders)
+        .where(
+            and(
+                eq(
+                    siteResourceAiProviders.siteResourceId,
+                    siteResource.siteResourceId
+                ),
+                eq(siteResourceAiProviders.accessMode, "select")
+            )
+        )
+        .limit(1);
+
+    if (!row) {
+        return "Set at least one attached AI provider to select mode before managing model lists";
+    }
+    return null;
+}
+
+/**
+ * Resource model entries must belong to select-mode attached providers, and
+ * listType must match the provider catalog entry (allow→allow, block→block).
+ */
+export async function assertPublicResourceModelEntriesValid(input: {
+    orgId: string;
+    resourceId: number;
+    models: ResourceAiModelEntry[];
+}): Promise<string | null> {
+    const uniqueModels = dedupeModelEntries(input.models);
+    if (uniqueModels.length === 0) {
+        return null;
+    }
+
+    const attachments = await db
+        .select({
+            providerId: resourceAiProviders.providerId,
+            accessMode: resourceAiProviders.accessMode,
+            enabled: resourceAiProviders.enabled
+        })
+        .from(resourceAiProviders)
+        .innerJoin(
+            aiProviders,
+            eq(resourceAiProviders.providerId, aiProviders.providerId)
+        )
+        .where(
+            and(
+                eq(resourceAiProviders.resourceId, input.resourceId),
+                eq(aiProviders.orgId, input.orgId)
+            )
+        );
+
+    return assertModelEntriesValid({
+        orgId: input.orgId,
+        modelEntries: uniqueModels,
+        attachments,
+        resourceLabel: "resource"
+    });
+}
+
+export async function assertSiteResourceModelEntriesValid(input: {
+    orgId: string;
+    siteResourceId: number;
+    models: ResourceAiModelEntry[];
+}): Promise<string | null> {
+    const uniqueModels = dedupeModelEntries(input.models);
+    if (uniqueModels.length === 0) {
+        return null;
+    }
+
+    const attachments = await db
+        .select({
+            providerId: siteResourceAiProviders.providerId,
+            accessMode: siteResourceAiProviders.accessMode,
+            enabled: siteResourceAiProviders.enabled
+        })
+        .from(siteResourceAiProviders)
+        .innerJoin(
+            aiProviders,
+            eq(siteResourceAiProviders.providerId, aiProviders.providerId)
+        )
+        .where(
+            and(
+                eq(
+                    siteResourceAiProviders.siteResourceId,
+                    input.siteResourceId
+                ),
+                eq(aiProviders.orgId, input.orgId)
+            )
+        );
+
+    return assertModelEntriesValid({
+        orgId: input.orgId,
+        modelEntries: uniqueModels,
+        attachments,
+        resourceLabel: "site resource"
+    });
+}
+
+function dedupeModelEntries(
+    models: ResourceAiModelEntry[]
+): ResourceAiModelEntry[] {
+    const byModelId = new Map(
+        models.map((m) => [m.modelId, m.listType] as const)
+    );
+    return [...byModelId.entries()].map(([modelId, listType]) => ({
+        modelId,
+        listType
+    }));
+}
+
+async function assertModelEntriesValid(input: {
+    orgId: string;
+    modelEntries: ResourceAiModelEntry[];
+    attachments: ResourceAiProviderAttachment[];
+    resourceLabel: string;
+}): Promise<string | null> {
+    const selectProviderIds = input.attachments
+        .filter((a) => a.accessMode === "select")
+        .map((a) => a.providerId);
+
+    if (selectProviderIds.length === 0) {
+        return "Set at least one attached AI provider to select mode before managing model lists";
+    }
+
+    const modelIds = input.modelEntries.map((m) => m.modelId);
+    const catalogRows = await db
+        .select({
+            modelId: aiModels.modelId,
+            modelKey: aiModels.modelKey,
+            listType: aiModels.listType,
+            providerId: aiModels.providerId,
+            enabled: aiModels.enabled
+        })
+        .from(aiModels)
+        .innerJoin(aiProviders, eq(aiModels.providerId, aiProviders.providerId))
+        .where(
+            and(
+                inArray(aiModels.modelId, modelIds),
+                inArray(aiModels.providerId, selectProviderIds),
+                eq(aiProviders.orgId, input.orgId)
+            )
+        );
+
+    if (catalogRows.length !== modelIds.length) {
+        return `One or more model IDs do not exist or do not belong to a select-mode provider on this ${input.resourceLabel}`;
+    }
+
+    const catalogById = new Map(catalogRows.map((row) => [row.modelId, row]));
+    const selectedAllowsByProvider = new Map<number, string[]>();
+    for (const entry of input.modelEntries) {
+        const catalog = catalogById.get(entry.modelId);
+        if (!catalog) {
+            return `One or more model IDs do not exist or do not belong to a select-mode provider on this ${input.resourceLabel}`;
+        }
+        if (catalog.listType !== entry.listType) {
+            return `Model ${entry.modelId} must use listType "${catalog.listType}" to match the provider catalog entry`;
+        }
+        if (!catalog.enabled) {
+            return `Model ${entry.modelId} is disabled on its provider`;
+        }
+        if (entry.listType === "allow") {
+            const keys = selectedAllowsByProvider.get(catalog.providerId) ?? [];
+            keys.push(catalog.modelKey);
+            selectedAllowsByProvider.set(catalog.providerId, keys);
+        }
+    }
+
+    const overlapError = await assertNoOverlappingModelKeys(input.attachments, {
+        selectedAllowsByProvider
+    });
+    if (overlapError) {
+        return overlapError.error;
+    }
+
+    return null;
+}

server/lib/aiModelCatalog.ts

@@ -0,0 +1,275 @@
+import fs from "node:fs";
+import axios from "axios";
+import config from "@server/lib/config";
+import logger from "@server/logger";
+import type { AiProviderType } from "@server/lib/aiProviderDefaults";
+
+export const CATALOG_PROVIDERS = [
+    "openai",
+    "anthropic",
+    "gemini",
+    "vertex",
+    "azure",
+    "bedrock"
+] as const;
+
+export type CatalogProvider = (typeof CATALOG_PROVIDERS)[number];
+
+const CATALOG_PROVIDER_SET = new Set<string>(CATALOG_PROVIDERS);
+
+// Each of our provider types maps to at most one catalog provider. Provider
+// types that proxy arbitrary underlying models (openRouter, vercelAiGateway,
+// custom) have no mapping.
+const PROVIDER_CATALOG_MAP: Record<
+    Exclude<AiProviderType, "custom">,
+    CatalogProvider | null
+> = {
+    openai: "openai",
+    anthropic: "anthropic",
+    googleGemini: "gemini",
+    vertexAi: "vertex",
+    bedrock: "bedrock",
+    microsoftFoundry: "azure",
+    openRouter: null,
+    vercelAiGateway: null
+};
+
+export function getCatalogProviderForType(
+    type: AiProviderType
+): CatalogProvider | null {
+    if (type === "custom") {
+        return null;
+    }
+    return PROVIDER_CATALOG_MAP[type];
+}
+
+export type AiModelCatalogEntry = {
+    provider: CatalogProvider;
+    model: string;
+    pricing: {
+        input: number | null;
+        output: number | null;
+        cacheRead: number | null;
+        reasoningOutput: number | null;
+    };
+};
+
+type RawCatalogEntry = {
+    id?: string;
+    name?: string;
+    model?: string;
+    provider: string;
+    input_cost_per_token?: number | null;
+    output_cost_per_token?: number | null;
+    cache_read_input_token_cost?: number | null;
+    output_cost_per_reasoning_token?: number | null;
+    pricing?: {
+        input?: number | null;
+        output?: number | null;
+        cacheRead?: number | null;
+        reasoningOutput?: number | null;
+    };
+};
+
+function normalizeCatalogProvider(raw: string): CatalogProvider | null {
+    if (CATALOG_PROVIDER_SET.has(raw)) {
+        return raw as CatalogProvider;
+    }
+    if (raw.startsWith("bedrock")) {
+        return "bedrock";
+    }
+    if (raw.startsWith("vertex")) {
+        return "vertex";
+    }
+    if (raw.startsWith("azure")) {
+        return "azure";
+    }
+    return null;
+}
+
+function normalizeEntry(raw: RawCatalogEntry): AiModelCatalogEntry | null {
+    const provider = normalizeCatalogProvider(raw.provider);
+    if (!provider) {
+        return null;
+    }
+
+    const model = raw.model ?? raw.name ?? raw.id;
+    if (!model) {
+        return null;
+    }
+
+    return {
+        provider,
+        model,
+        pricing: {
+            input: raw.pricing?.input ?? raw.input_cost_per_token ?? null,
+            output: raw.pricing?.output ?? raw.output_cost_per_token ?? null,
+            cacheRead:
+                raw.pricing?.cacheRead ??
+                raw.cache_read_input_token_cost ??
+                null,
+            reasoningOutput:
+                raw.pricing?.reasoningOutput ??
+                raw.output_cost_per_reasoning_token ??
+                null
+        }
+    };
+}
+
+function providerKey(provider: CatalogProvider, key: string): string {
+    return `${provider}\0${key}`;
+}
+
+export class AiModelCatalog {
+    private entries: AiModelCatalogEntry[] = [];
+    private byProvider = new Map<CatalogProvider, AiModelCatalogEntry[]>();
+    private byProviderAndKey = new Map<string, AiModelCatalogEntry>();
+    private byKey = new Map<string, AiModelCatalogEntry[]>();
+    private refreshTimer: NodeJS.Timeout | null = null;
+
+    /**
+     * Loads the catalog into memory and schedules periodic background refreshes.
+     * Call once at server startup.
+     */
+    async init(): Promise<void> {
+        await this.refresh();
+        this.scheduleNextRefresh();
+    }
+
+    /** Exact lookup by catalog provider and model key. */
+    get(
+        provider: CatalogProvider,
+        key: string
+    ): AiModelCatalogEntry | undefined {
+        return this.byProviderAndKey.get(providerKey(provider, key));
+    }
+
+    /** All models for a catalog provider. */
+    list(provider: CatalogProvider): AiModelCatalogEntry[] {
+        return this.byProvider.get(provider) ?? [];
+    }
+
+    /** All catalog entries that share a model key, across providers. */
+    listByKey(key: string): AiModelCatalogEntry[] {
+        return this.byKey.get(key) ?? [];
+    }
+
+    /** Full in-memory catalog. */
+    getAll(): AiModelCatalogEntry[] {
+        return this.entries;
+    }
+
+    private setEntries(entries: AiModelCatalogEntry[]): void {
+        const byProvider = new Map<CatalogProvider, AiModelCatalogEntry[]>();
+        const byProviderAndKey = new Map<string, AiModelCatalogEntry>();
+        const byKey = new Map<string, AiModelCatalogEntry[]>();
+
+        for (const entry of entries) {
+            const list = byProvider.get(entry.provider) ?? [];
+            list.push(entry);
+            byProvider.set(entry.provider, list);
+
+            const mapKey = providerKey(entry.provider, entry.model);
+            if (!byProviderAndKey.has(mapKey)) {
+                byProviderAndKey.set(mapKey, entry);
+            }
+
+            const keyList = byKey.get(entry.model) ?? [];
+            keyList.push(entry);
+            byKey.set(entry.model, keyList);
+        }
+
+        this.entries = entries;
+        this.byProvider = byProvider;
+        this.byProviderAndKey = byProviderAndKey;
+        this.byKey = byKey;
+    }
+
+    private async fetchFromFile(
+        filePath: string
+    ): Promise<AiModelCatalogEntry[] | null> {
+        try {
+            if (!fs.existsSync(filePath)) {
+                logger.warn(
+                    `AI model catalog file not found at ${filePath}; cost calculation will fall back to unknown pricing`
+                );
+                return null;
+            }
+            const raw = fs.readFileSync(filePath, "utf-8");
+            const parsed = JSON.parse(raw) as { data: RawCatalogEntry[] };
+            return (parsed.data ?? [])
+                .map(normalizeEntry)
+                .filter((e): e is AiModelCatalogEntry => e != null);
+        } catch (error) {
+            logger.warn("Failed to read AI model catalog file", { error });
+            return null;
+        }
+    }
+
+    private async fetchFromUpstream(
+        upstreamUrl: string
+    ): Promise<AiModelCatalogEntry[] | null> {
+        try {
+            const res = await axios.get<{ data: RawCatalogEntry[] }>(
+                upstreamUrl,
+                { timeout: 15_000 }
+            );
+            return (res.data?.data ?? [])
+                .map(normalizeEntry)
+                .filter((e): e is AiModelCatalogEntry => e != null);
+        } catch (error: any) {
+            logger.warn(
+                `Failed to fetch AI model catalog from ${upstreamUrl}: ${error.message || error}`
+            );
+            return null;
+        }
+    }
+
+    private async refresh(): Promise<void> {
+        const { file, upstream_url } = config.getRawConfig().ai.model_catalog;
+
+        const fetched = file
+            ? await this.fetchFromFile(file)
+            : await this.fetchFromUpstream(upstream_url);
+
+        if (fetched) {
+            this.setEntries(fetched);
+            logger.debug(
+                `AI model catalog refreshed: ${this.entries.length} models loaded`
+            );
+        } else {
+            logger.debug(
+                "AI model catalog refresh failed; keeping previously loaded catalog in memory"
+            );
+        }
+    }
+
+    private scheduleNextRefresh(): void {
+        const { refresh_interval_min_hours, refresh_interval_max_hours } =
+            config.getRawConfig().ai.model_catalog;
+
+        // Jittered rather than fixed so that many self-hosted instances don't
+        // all hit the upstream catalog endpoint at the same moment.
+        const minMs = refresh_interval_min_hours * 60 * 60 * 1000;
+        const maxMs = refresh_interval_max_hours * 60 * 60 * 1000;
+        const delayMs = minMs + Math.random() * Math.max(0, maxMs - minMs);
+
+        if (this.refreshTimer) {
+            clearTimeout(this.refreshTimer);
+        }
+        this.refreshTimer = setTimeout(async () => {
+            await this.refresh();
+            this.scheduleNextRefresh();
+        }, delayMs);
+    }
+}
+
+export const aiModelCatalog = new AiModelCatalog();
+
+/**
+ * Loads the AI model pricing catalog into memory and schedules periodic
+ * background refreshes. Call once at server startup.
+ */
+export async function initAiModelCatalog(): Promise<void> {
+    await aiModelCatalog.init();
+}

server/lib/aiModelKeyMatch.ts

@@ -0,0 +1,124 @@
+const modelKeyRegexCache = new Map<string, RegExp>();
+
+export function isModelKeyPattern(key: string): boolean {
+    return key.includes("*") || key.includes("?");
+}
+
+function getModelKeyRegex(pattern: string): RegExp {
+    let regex = modelKeyRegexCache.get(pattern);
+    if (!regex) {
+        const escaped = pattern.replace(/[.+^${}()|[\]\\]/g, "\\$&");
+        regex = new RegExp(
+            `^${escaped.replace(/\*/g, ".*").replace(/\?/g, ".")}$`
+        );
+        modelKeyRegexCache.set(pattern, regex);
+    }
+    return regex;
+}
+
+export function modelKeyMatches(
+    pattern: string,
+    requestedModel: string
+): boolean {
+    return getModelKeyRegex(pattern).test(requestedModel);
+}
+
+function wildcardCharCount(key: string): number {
+    let count = 0;
+    for (const char of key) {
+        if (char === "*" || char === "?") {
+            count += 1;
+        }
+    }
+    return count;
+}
+
+function literalLength(key: string): number {
+    return key.replace(/[*?]/g, "").length;
+}
+
+/**
+ * Sort comparator: more specific patterns sort before less specific ones
+ * (negative when `a` is more specific than `b`).
+ *
+ * 1. Exact keys beat patterns
+ * 2. Fewer wildcard characters win
+ * 3. Longer literal length wins
+ */
+export function compareModelKeySpecificity(a: string, b: string): number {
+    const aIsPattern = isModelKeyPattern(a);
+    const bIsPattern = isModelKeyPattern(b);
+
+    if (aIsPattern !== bIsPattern) {
+        return aIsPattern ? 1 : -1;
+    }
+
+    const wildcardDiff = wildcardCharCount(a) - wildcardCharCount(b);
+    if (wildcardDiff !== 0) {
+        return wildcardDiff;
+    }
+
+    return literalLength(b) - literalLength(a);
+}
+
+/**
+ * Attach-time conflict check. Detects identical keys and exact-vs-pattern
+ * matches. Does not attempt full glob intersection.
+ */
+export function modelKeysConflict(a: string, b: string): boolean {
+    if (a === b) {
+        return true;
+    }
+
+    const aIsPattern = isModelKeyPattern(a);
+    const bIsPattern = isModelKeyPattern(b);
+
+    if (aIsPattern === bIsPattern) {
+        return false;
+    }
+
+    if (aIsPattern) {
+        return modelKeyMatches(a, b);
+    }
+
+    return modelKeyMatches(b, a);
+}
+
+/**
+ * Provider-layer policy: empty allowlist denies all. Blocklist only applies
+ * after an allow match.
+ */
+export function isAllowedByLists(
+    requested: string,
+    allows: string[],
+    blocks: string[]
+): boolean {
+    if (allows.length === 0) {
+        return false;
+    }
+    if (!allows.some((pattern) => modelKeyMatches(pattern, requested))) {
+        return false;
+    }
+    if (blocks.some((pattern) => modelKeyMatches(pattern, requested))) {
+        return false;
+    }
+    return true;
+}
+
+/**
+ * Among allow patterns that match `requested`, return the most specific one,
+ * or null if none match.
+ */
+export function mostSpecificMatchingAllow(
+    requested: string,
+    allows: string[]
+): string | null {
+    const matching = allows.filter((pattern) =>
+        modelKeyMatches(pattern, requested)
+    );
+    if (matching.length === 0) {
+        return null;
+    }
+    matching.sort(compareModelKeySpecificity);
+    return matching[0];
+}

server/lib/aiModelPricing.ts

@@ -0,0 +1,148 @@
+import type { AiProviderType } from "@server/lib/aiProviderDefaults";
+import type { AiUsage } from "@server/lib/aiUsageExtraction";
+import {
+    aiModelCatalog,
+    getCatalogProviderForType,
+    type AiModelCatalogEntry,
+    type CatalogProvider
+} from "@server/lib/aiModelCatalog";
+
+export type AiModelPricing = {
+    inputCostPerToken: number | null;
+    outputCostPerToken: number | null;
+    cacheReadInputTokenCost: number | null;
+    outputCostPerReasoningToken: number | null;
+    // True when the match came from a different catalog provider than the
+    // one mapped to this provider's type (e.g. an openRouter/custom model
+    // id that only matched a global search across every provider). Costs
+    // found this way are a best-effort approximation, not a guarantee the
+    // upstream provider bills at the same rate.
+    approximate: boolean;
+};
+
+function stripVendorPrefix(modelId: string): string | null {
+    const idx = modelId.indexOf("/");
+    if (idx === -1 || idx === modelId.length - 1) {
+        return null;
+    }
+    return modelId.slice(idx + 1);
+}
+
+function toPricing(
+    entry: AiModelCatalogEntry,
+    approximate: boolean
+): AiModelPricing {
+    return {
+        inputCostPerToken: entry.pricing.input,
+        outputCostPerToken: entry.pricing.output,
+        cacheReadInputTokenCost: entry.pricing.cacheRead,
+        outputCostPerReasoningToken: entry.pricing.reasoningOutput,
+        approximate
+    };
+}
+
+function findEntry(
+    modelId: string,
+    provider: CatalogProvider | null
+): AiModelCatalogEntry | null {
+    const candidates = [modelId, stripVendorPrefix(modelId)].filter(
+        (v): v is string => v != null
+    );
+
+    for (const key of candidates) {
+        if (provider) {
+            const match = aiModelCatalog.get(provider, key);
+            if (match) {
+                return match;
+            }
+            continue;
+        }
+
+        const match = aiModelCatalog.listByKey(key)[0];
+        if (match) {
+            return match;
+        }
+    }
+    return null;
+}
+
+/**
+ * Looks up per-token pricing for a model, scoped first to the catalog
+ * provider that corresponds to our provider type, then falling back to a
+ * global search across every provider (marked `approximate`) for provider
+ * types that proxy arbitrary underlying models.
+ */
+export function getModelPricing(
+    providerType: AiProviderType,
+    modelId: string | undefined
+): AiModelPricing | null {
+    if (!modelId) {
+        return null;
+    }
+
+    const catalogProvider = getCatalogProviderForType(providerType);
+
+    if (catalogProvider) {
+        const scoped = findEntry(modelId, catalogProvider);
+        if (scoped) {
+            return toPricing(scoped, false);
+        }
+    }
+
+    const fallback = findEntry(modelId, null);
+    if (fallback) {
+        return toPricing(fallback, true);
+    }
+
+    return null;
+}
+
+export type AiCostBreakdown = {
+    promptCost: number;
+    cacheReadCost: number;
+    cacheWriteCost: number;
+    completionCost: number;
+    reasoningCost: number;
+    totalCost: number;
+};
+
+/**
+ * Computes a $ cost breakdown for a usage record given a model's pricing.
+ * Cache writes and reasoning tokens fall back to the normal input/output
+ * rate respectively when the catalog has no dedicated rate for them (the
+ * catalog has no cache-write field at all, and only some models report a
+ * distinct reasoning rate).
+ */
+export function calculateAiCost(
+    pricing: AiModelPricing | null,
+    usage: AiUsage
+): AiCostBreakdown | null {
+    if (!pricing) {
+        return null;
+    }
+
+    const inputRate = pricing.inputCostPerToken ?? 0;
+    const outputRate = pricing.outputCostPerToken ?? 0;
+    const cacheReadRate = pricing.cacheReadInputTokenCost ?? inputRate;
+    const reasoningRate = pricing.outputCostPerReasoningToken ?? outputRate;
+
+    const promptCost = usage.promptTokens * inputRate;
+    const cacheReadCost = usage.cacheReadTokens * cacheReadRate;
+    const cacheWriteCost = usage.cacheWriteTokens * inputRate;
+    const completionCost = usage.completionTokens * outputRate;
+    const reasoningCost = usage.reasoningTokens * reasoningRate;
+
+    return {
+        promptCost,
+        cacheReadCost,
+        cacheWriteCost,
+        completionCost,
+        reasoningCost,
+        totalCost:
+            promptCost +
+            cacheReadCost +
+            cacheWriteCost +
+            completionCost +
+            reasoningCost
+    };
+}

server/lib/aiProviderDefaults.ts

@@ -0,0 +1,262 @@
+import { decrypt, encrypt } from "@server/lib/crypto";
+import {
+    parseCapabilities,
+    type AiCapability
+} from "@server/lib/aiCapabilities";
+
+export type AiProviderType =
+    | "openai"
+    | "anthropic"
+    | "googleGemini"
+    | "vertexAi"
+    | "bedrock"
+    | "microsoftFoundry"
+    | "openRouter"
+    | "vercelAiGateway"
+    | "custom";
+
+export const AI_PROVIDER_AUTH_TYPES = [
+    "bearer",
+    "x-api-key",
+    "x-goog-api-key",
+    "hec",
+    "cf-aig-authorization",
+    "none",
+    "passthrough"
+] as const;
+
+export type AiProviderAuthType = (typeof AI_PROVIDER_AUTH_TYPES)[number];
+export type AiBudgetUnit = "usd" | "tokens";
+export type AiProviderRoutingMode = "url" | "target";
+
+type AiProviderDefaults = {
+    upstreamUrl: string | null;
+    authType: AiProviderAuthType;
+    capabilities: readonly AiCapability[];
+};
+
+export const AI_PROVIDER_DEFAULTS: Record<
+    Exclude<AiProviderType, "custom">,
+    AiProviderDefaults
+> = {
+    openai: {
+        upstreamUrl: "https://api.openai.com/v1",
+        authType: "bearer",
+        capabilities: ["openai_chat", "openai_responses"]
+    },
+    anthropic: {
+        upstreamUrl: "https://api.anthropic.com",
+        authType: "x-api-key",
+        capabilities: ["anthropic_messages"]
+    },
+    googleGemini: {
+        upstreamUrl: "https://generativelanguage.googleapis.com",
+        authType: "x-goog-api-key",
+        capabilities: ["gemini_generate_content"]
+    },
+    vertexAi: {
+        upstreamUrl: null,
+        authType: "bearer",
+        capabilities: ["google_generate_content", "google_raw_predict"]
+    },
+    bedrock: {
+        upstreamUrl: "https://bedrock-runtime.us-east-1.amazonaws.com",
+        authType: "bearer",
+        capabilities: ["bedrock_converse"]
+    },
+    microsoftFoundry: {
+        upstreamUrl: null,
+        authType: "bearer",
+        capabilities: ["openai_chat", "openai_responses", "anthropic_messages"]
+    },
+    openRouter: {
+        upstreamUrl: "https://openrouter.ai/api/v1",
+        authType: "bearer",
+        capabilities: ["openai_chat"]
+    },
+    vercelAiGateway: {
+        upstreamUrl: "https://ai-gateway.vercel.sh/v1",
+        authType: "bearer",
+        capabilities: ["openai_chat", "openai_responses"]
+    }
+};
+
+const CONFLICTING_AUTH_HEADERS = [
+    "authorization",
+    "x-api-key",
+    "x-goog-api-key",
+    "cf-aig-authorization"
+] as const;
+
+export function authTypeRequiresApiKey(authType: AiProviderAuthType): boolean {
+    return authType !== "none" && authType !== "passthrough";
+}
+
+export function providerRequiresUpstreamUrl(
+    type: AiProviderType,
+    routingMode: AiProviderRoutingMode = "url"
+): boolean {
+    if (routingMode === "target") {
+        return false;
+    }
+    if (type === "custom") {
+        return true;
+    }
+    return AI_PROVIDER_DEFAULTS[type].upstreamUrl === null;
+}
+
+export function resolveAiProviderCreateFields(input: {
+    type: AiProviderType;
+    upstreamUrl?: string | null;
+    authType?: AiProviderAuthType | null;
+    routingMode?: AiProviderRoutingMode | null;
+}): {
+    upstreamUrl: string | null;
+    authType: AiProviderAuthType;
+    routingMode: AiProviderRoutingMode;
+} {
+    const routingMode =
+        input.type === "custom" ? (input.routingMode ?? "url") : "url";
+
+    if (routingMode === "target") {
+        return {
+            upstreamUrl: null,
+            authType: input.authType ?? "bearer",
+            routingMode
+        };
+    }
+
+    if (input.type === "custom") {
+        return {
+            upstreamUrl: input.upstreamUrl ?? null,
+            authType: input.authType ?? "bearer",
+            routingMode
+        };
+    }
+
+    const defaults = AI_PROVIDER_DEFAULTS[input.type];
+    return {
+        upstreamUrl: input.upstreamUrl ?? defaults.upstreamUrl,
+        authType: input.authType ?? defaults.authType,
+        routingMode
+    };
+}
+
+export type AiProviderHeader = { name: string; value: string };
+
+export function serializeAiProviderHeaders(
+    headers: AiProviderHeader[] | null | undefined,
+    secret: string
+): string | null {
+    if (!headers || headers.length === 0) {
+        return null;
+    }
+    return encrypt(JSON.stringify(headers), secret);
+}
+
+export function parseAiProviderHeaders(
+    raw: string | null | undefined,
+    secret: string
+): AiProviderHeader[] {
+    if (!raw) {
+        return [];
+    }
+    try {
+        const decrypted = decrypt(raw, secret);
+        const parsed = JSON.parse(decrypted);
+        if (!Array.isArray(parsed)) {
+            return [];
+        }
+        return parsed.filter(
+            (h): h is AiProviderHeader =>
+                h != null &&
+                typeof h === "object" &&
+                typeof h.name === "string" &&
+                typeof h.value === "string"
+        );
+    } catch {
+        return [];
+    }
+}
+
+export function applyAiProviderCustomHeaders(
+    headers: Record<string, string>,
+    raw: string | null | undefined,
+    secret: string
+): void {
+    for (const { name, value } of parseAiProviderHeaders(raw, secret)) {
+        headers[name] = value;
+    }
+}
+
+/**
+ * Apply provider auth to upstream headers.
+ * - Injected modes: strip client auth headers, then set the provider key.
+ * - none: strip client auth headers, send no auth.
+ * - passthrough: leave client auth headers as-is.
+ */
+export function applyAiProviderAuthHeaders(
+    headers: Record<string, string>,
+    authType: AiProviderAuthType,
+    apiKey: string | null
+): void {
+    if (authType === "passthrough") {
+        return;
+    }
+
+    for (const name of CONFLICTING_AUTH_HEADERS) {
+        for (const key of Object.keys(headers)) {
+            if (key.toLowerCase() === name) {
+                delete headers[key];
+            }
+        }
+    }
+
+    if (authType === "none") {
+        return;
+    }
+
+    if (!apiKey) {
+        throw new Error(`API key required for authType ${authType}`);
+    }
+
+    switch (authType) {
+        case "bearer":
+            headers["Authorization"] = `Bearer ${apiKey}`;
+            break;
+        case "x-api-key":
+            headers["x-api-key"] = apiKey;
+            break;
+        case "x-goog-api-key":
+            headers["x-goog-api-key"] = apiKey;
+            break;
+        case "hec":
+            headers["Authorization"] = `Splunk ${apiKey}`;
+            break;
+        case "cf-aig-authorization":
+            headers["cf-aig-authorization"] = `Bearer ${apiKey}`;
+            break;
+    }
+}
+
+export function resolveCapabilitiesForCreate(input: {
+    type: AiProviderType;
+    capabilities?: AiCapability[] | null;
+}): AiCapability[] {
+    if (input.capabilities != null) {
+        return parseCapabilities(input.capabilities);
+    }
+    if (input.type === "custom") {
+        return [];
+    }
+    return [...AI_PROVIDER_DEFAULTS[input.type].capabilities];
+}
+
+export function defaultsForProviderType(
+    type: AiProviderType
+): readonly AiCapability[] {
+    if (type === "custom") {
+        return [];
+    }
+    return AI_PROVIDER_DEFAULTS[type].capabilities;
+}

server/lib/aiUsageExtraction.ts

@@ -0,0 +1,468 @@
+import { encode } from "gpt-tokenizer";
+import type { AiCapability } from "@server/lib/aiCapabilities";
+import logger from "@server/logger";
+
+export type AiUsage = {
+    // Input tokens billed at the normal input rate (i.e. NOT already
+    // covered by cacheReadTokens/cacheWriteTokens below).
+    promptTokens: number;
+    cacheReadTokens: number;
+    cacheWriteTokens: number;
+    // Output tokens billed at the normal output rate (i.e. NOT already
+    // covered by reasoningTokens below).
+    completionTokens: number;
+    reasoningTokens: number;
+    // True when these numbers are our own best-guess estimate (the upstream
+    // response didn't report usage), rather than provider-reported figures.
+    estimated: boolean;
+};
+
+function emptyUsage(): AiUsage {
+    return {
+        promptTokens: 0,
+        cacheReadTokens: 0,
+        cacheWriteTokens: 0,
+        completionTokens: 0,
+        reasoningTokens: 0,
+        estimated: false
+    };
+}
+
+/**
+ * Scans raw (possibly binary-framed, e.g. Bedrock's vnd.amazon.eventstream)
+ * text for `"fieldName":123` occurrences and returns the last value seen for
+ * each field. Used as a best-effort fallback for response shapes we can't
+ * fully parse as JSON/SSE (streaming Bedrock, raw predict passthroughs).
+ */
+function scanNumericFields(
+    text: string,
+    fields: string[]
+): Record<string, number> {
+    const out: Record<string, number> = {};
+    for (const field of fields) {
+        const re = new RegExp(`"${field}"\\s*:\\s*(\\d+)`, "g");
+        let match: RegExpExecArray | null;
+        while ((match = re.exec(text)) !== null) {
+            out[field] = Number(match[1]);
+        }
+    }
+    return out;
+}
+
+function sseDataFrames(text: string): string[] {
+    const frames: string[] = [];
+    for (const rawFrame of text.split(/\r?\n\r?\n/)) {
+        for (const line of rawFrame.split(/\r?\n/)) {
+            if (!line.startsWith("data:")) continue;
+            const data = line.slice("data:".length).trim();
+            if (data && data !== "[DONE]") {
+                frames.push(data);
+            }
+        }
+    }
+    return frames;
+}
+
+function tryParseJson(text: string): any | null {
+    try {
+        return JSON.parse(text);
+    } catch {
+        return null;
+    }
+}
+
+function extractOpenAiChat(text: string, isStream: boolean): AiUsage | null {
+    let usage: any = null;
+
+    if (isStream) {
+        for (const frame of sseDataFrames(text)) {
+            const parsed = tryParseJson(frame);
+            if (parsed?.usage) {
+                usage = parsed.usage;
+            }
+        }
+    } else {
+        usage = tryParseJson(text)?.usage ?? null;
+    }
+
+    if (!usage) {
+        return null;
+    }
+
+    const cacheReadTokens = usage.prompt_tokens_details?.cached_tokens ?? 0;
+    const reasoningTokens =
+        usage.completion_tokens_details?.reasoning_tokens ?? 0;
+
+    return {
+        promptTokens: Math.max(0, (usage.prompt_tokens ?? 0) - cacheReadTokens),
+        cacheReadTokens,
+        cacheWriteTokens: 0,
+        completionTokens: Math.max(
+            0,
+            (usage.completion_tokens ?? 0) - reasoningTokens
+        ),
+        reasoningTokens,
+        estimated: false
+    };
+}
+
+function extractOpenAiResponses(
+    text: string,
+    isStream: boolean
+): AiUsage | null {
+    let usage: any = null;
+
+    if (isStream) {
+        for (const frame of sseDataFrames(text)) {
+            const parsed = tryParseJson(frame);
+            if (parsed?.type === "response.completed" && parsed?.response?.usage) {
+                usage = parsed.response.usage;
+            } else if (parsed?.usage) {
+                usage = parsed.usage;
+            }
+        }
+    } else {
+        const parsed = tryParseJson(text);
+        usage = parsed?.usage ?? parsed?.response?.usage ?? null;
+    }
+
+    if (!usage) {
+        return null;
+    }
+
+    const cacheReadTokens = usage.input_tokens_details?.cached_tokens ?? 0;
+    const reasoningTokens = usage.output_tokens_details?.reasoning_tokens ?? 0;
+
+    return {
+        promptTokens: Math.max(0, (usage.input_tokens ?? 0) - cacheReadTokens),
+        cacheReadTokens,
+        cacheWriteTokens: 0,
+        completionTokens: Math.max(
+            0,
+            (usage.output_tokens ?? 0) - reasoningTokens
+        ),
+        reasoningTokens,
+        estimated: false
+    };
+}
+
+function extractAnthropicMessages(
+    text: string,
+    isStream: boolean
+): AiUsage | null {
+    let inputTokens = 0;
+    let cacheReadTokens = 0;
+    let cacheWriteTokens = 0;
+    let outputTokens = 0;
+    let found = false;
+
+    const applyUsage = (usage: any) => {
+        if (!usage) return;
+        found = true;
+        if (typeof usage.input_tokens === "number") {
+            inputTokens = usage.input_tokens;
+        }
+        if (typeof usage.cache_read_input_tokens === "number") {
+            cacheReadTokens = usage.cache_read_input_tokens;
+        }
+        if (typeof usage.cache_creation_input_tokens === "number") {
+            cacheWriteTokens = usage.cache_creation_input_tokens;
+        }
+        if (typeof usage.output_tokens === "number") {
+            outputTokens = usage.output_tokens;
+        }
+    };
+
+    if (isStream) {
+        for (const frame of sseDataFrames(text)) {
+            const parsed = tryParseJson(frame);
+            if (!parsed) continue;
+            applyUsage(parsed.message?.usage);
+            applyUsage(parsed.usage);
+        }
+    } else {
+        applyUsage(tryParseJson(text)?.usage);
+    }
+
+    if (!found) {
+        return null;
+    }
+
+    return {
+        promptTokens: inputTokens,
+        cacheReadTokens,
+        cacheWriteTokens,
+        completionTokens: outputTokens,
+        // Anthropic bills extended-thinking output at the normal output
+        // rate, so there's no separate reasoning bucket to report.
+        reasoningTokens: 0,
+        estimated: false
+    };
+}
+
+function extractGoogleGenerateContent(
+    text: string,
+    _isStream: boolean
+): AiUsage | null {
+    // Both the plain-JSON-array stream format and the SSE (?alt=sse) format
+    // repeat a cumulative `usageMetadata` object per chunk; the regex scan
+    // below naturally picks up the last (most complete) one either way.
+    const fields = scanNumericFields(text, [
+        "promptTokenCount",
+        "candidatesTokenCount",
+        "cachedContentTokenCount",
+        "thoughtsTokenCount"
+    ]);
+
+    if (fields.promptTokenCount === undefined) {
+        return null;
+    }
+
+    const cacheReadTokens = fields.cachedContentTokenCount ?? 0;
+    const reasoningTokens = fields.thoughtsTokenCount ?? 0;
+
+    return {
+        promptTokens: Math.max(0, fields.promptTokenCount - cacheReadTokens),
+        cacheReadTokens,
+        cacheWriteTokens: 0,
+        completionTokens: fields.candidatesTokenCount ?? 0,
+        reasoningTokens,
+        estimated: false
+    };
+}
+
+function extractBedrockConverse(
+    text: string,
+    _isStream: boolean
+): AiUsage | null {
+    // Non-streaming responses are plain JSON; converse-stream frames the
+    // final `metadata` event's usage object inside binary event-stream
+    // framing, but the JSON text survives intact inside that binary
+    // envelope, so the same field scan works for both.
+    const parsed = tryParseJson(text);
+    const usage = parsed?.usage;
+    if (usage) {
+        const cacheReadTokens = usage.cacheReadInputTokens ?? 0;
+        return {
+            promptTokens: Math.max(0, (usage.inputTokens ?? 0) - cacheReadTokens),
+            cacheReadTokens,
+            cacheWriteTokens: usage.cacheWriteInputTokens ?? 0,
+            completionTokens: usage.outputTokens ?? 0,
+            reasoningTokens: 0,
+            estimated: false
+        };
+    }
+
+    const fields = scanNumericFields(text, [
+        "inputTokens",
+        "outputTokens",
+        "cacheReadInputTokens",
+        "cacheWriteInputTokens"
+    ]);
+    if (fields.inputTokens === undefined) {
+        return null;
+    }
+    const cacheReadTokens = fields.cacheReadInputTokens ?? 0;
+    return {
+        promptTokens: Math.max(0, fields.inputTokens - cacheReadTokens),
+        cacheReadTokens,
+        cacheWriteTokens: fields.cacheWriteInputTokens ?? 0,
+        completionTokens: fields.outputTokens ?? 0,
+        reasoningTokens: 0,
+        estimated: false
+    };
+}
+
+function extractBedrockModelInvoke(
+    text: string,
+    _isStream: boolean,
+    headers: Headers
+): AiUsage | null {
+    // Non-streaming invoke reports counts via response headers regardless
+    // of the underlying model's payload format.
+    const headerInput = headers.get("x-amzn-bedrock-input-token-count");
+    const headerOutput = headers.get("x-amzn-bedrock-output-token-count");
+    if (headerInput !== null || headerOutput !== null) {
+        return {
+            promptTokens: Number(headerInput ?? 0),
+            cacheReadTokens: 0,
+            cacheWriteTokens: 0,
+            completionTokens: Number(headerOutput ?? 0),
+            reasoningTokens: 0,
+            estimated: false
+        };
+    }
+
+    // invoke-with-response-stream has no equivalent headers; the model's
+    // own usage shape (frequently Anthropic-style on Bedrock) is embedded
+    // inside binary event-stream framing, so fall back to a couple of
+    // known field-name shapes via regex.
+    const anthropicStyle = extractAnthropicMessages(text, true);
+    if (anthropicStyle) {
+        return anthropicStyle;
+    }
+
+    const fields = scanNumericFields(text, [
+        "inputTokenCount",
+        "outputTokenCount"
+    ]);
+    if (fields.inputTokenCount === undefined) {
+        return null;
+    }
+    return {
+        promptTokens: fields.inputTokenCount,
+        cacheReadTokens: 0,
+        cacheWriteTokens: 0,
+        completionTokens: fields.outputTokenCount ?? 0,
+        reasoningTokens: 0,
+        estimated: false
+    };
+}
+
+const EXTRACTORS: Record<
+    AiCapability,
+    (text: string, isStream: boolean, headers: Headers) => AiUsage | null
+> = {
+    openai_chat: extractOpenAiChat,
+    openai_responses: extractOpenAiResponses,
+    anthropic_messages: extractAnthropicMessages,
+    gemini_generate_content: extractGoogleGenerateContent,
+    google_generate_content: extractGoogleGenerateContent,
+    // rawPredict is a passthrough to whatever the underlying publisher
+    // model speaks (often Anthropic-shaped on Vertex); try that, then give
+    // up to the token-count estimate.
+    google_raw_predict: (text, isStream) =>
+        extractAnthropicMessages(text, isStream),
+    bedrock_model_invoke: extractBedrockModelInvoke,
+    bedrock_converse: extractBedrockConverse
+};
+
+/**
+ * Attempts to pull provider-reported token usage out of an upstream AI
+ * gateway response. Returns null if the response didn't contain (or we
+ * couldn't find) usage data, in which case callers should fall back to
+ * `estimateUsage`.
+ */
+export function extractUsage(
+    capability: AiCapability,
+    responseText: string,
+    isStream: boolean,
+    headers: Headers
+): AiUsage | null {
+    try {
+        return EXTRACTORS[capability](responseText, isStream, headers);
+    } catch (error) {
+        logger.debug("Failed to extract AI usage from response", {
+            capability,
+            error
+        });
+        return null;
+    }
+}
+
+/**
+ * Best-guess token estimate for when the provider doesn't report usage.
+ * Uses OpenAI's BPE tokenizer as a stand-in for whatever tokenizer the
+ * actual model uses - close enough for an approximate cost figure, not
+ * exact for non-OpenAI models.
+ */
+export function estimateUsage(
+    promptText: string,
+    completionText: string
+): AiUsage {
+    const usage = emptyUsage();
+    usage.estimated = true;
+    try {
+        usage.promptTokens = promptText ? encode(promptText).length : 0;
+    } catch (error) {
+        logger.debug("Failed to estimate prompt tokens", { error });
+    }
+    try {
+        usage.completionTokens = completionText
+            ? encode(completionText).length
+            : 0;
+    } catch (error) {
+        logger.debug("Failed to estimate completion tokens", { error });
+    }
+    return usage;
+}
+
+/**
+ * OpenAI's Chat Completions API only includes a `usage` field in a
+ * streaming response when the request opts in via `stream_options:
+ * {include_usage: true}` - unlike the Responses API, Anthropic, Gemini and
+ * Bedrock, which report usage in a streaming response by default. Returns
+ * whether we need to inject that option ourselves to be able to track cost.
+ */
+export function needsStreamUsageInjection(
+    capability: AiCapability,
+    body: any
+): boolean {
+    return (
+        capability === "openai_chat" &&
+        body?.stream === true &&
+        body?.stream_options?.include_usage !== true
+    );
+}
+
+/**
+ * Returns a shallow-cloned body with `stream_options.include_usage`
+ * injected, for capabilities/requests where `needsStreamUsageInjection`
+ * is true. Leaves the original body untouched.
+ */
+export function withStreamUsageOption(body: any): any {
+    return {
+        ...body,
+        stream_options: { ...body.stream_options, include_usage: true }
+    };
+}
+
+/**
+ * When we injected stream_options.include_usage ourselves (the caller
+ * didn't ask for it), OpenAI appends an extra terminal SSE frame with an
+ * empty `choices: []` array carrying only the usage data. Callers that
+ * don't expect that shape (most minimal SSE parsers assume a non-empty
+ * choices array) shouldn't see it, so it's stripped back out of the bytes
+ * forwarded to the client.
+ */
+export function stripInjectedUsageFrame(sseText: string): string {
+    const parts = sseText.split(/(\r?\n\r?\n)/);
+    let out = "";
+    for (let i = 0; i < parts.length; i += 2) {
+        const frame = parts[i];
+        const separator = parts[i + 1] ?? "";
+        const dataLine = frame
+            .split(/\r?\n/)
+            .find((line) => line.startsWith("data:"));
+        if (dataLine) {
+            const data = dataLine.slice("data:".length).trim();
+            const parsed = data !== "[DONE]" ? tryParseJson(data) : null;
+            if (parsed && Array.isArray(parsed.choices) && parsed.choices.length === 0 && parsed.usage) {
+                continue;
+            }
+        }
+        out += frame + separator;
+    }
+    return out;
+}
+
+/**
+ * Best-effort extraction of the model the upstream provider actually
+ * served, which some gateways/routers echo back and which may differ from
+ * the model the caller requested (e.g. an alias resolving to a dated
+ * snapshot). Falls back to the caller's requested model when absent.
+ */
+export function extractResponseModel(responseText: string): string | null {
+    const match = responseText.match(/"model"\s*:\s*"([^"]+)"/);
+    return match ? match[1] : null;
+}
+
+export function isUsageEmpty(usage: AiUsage): boolean {
+    return (
+        usage.promptTokens === 0 &&
+        usage.cacheReadTokens === 0 &&
+        usage.cacheWriteTokens === 0 &&
+        usage.completionTokens === 0 &&
+        usage.reasoningTokens === 0
+    );
+}

server/lib/alerts/events/healthCheckEvents.ts

@@ -202,6 +202,10 @@ async function handleResource(
         return;
     }
 
+    if (!target.resourceId) {
+        return;
+    }
+
     const [resource] = await trx
         .select()
         .from(resources)
@@ -227,9 +231,7 @@ async function handleResource(
 
     let health = "healthy";
     const allUnknown = monitoredTargets.length === 0;
-    const allHealthy = monitoredTargets.every(
-        (t) => t.hcHealth === "healthy"
-    );
+    const allHealthy = monitoredTargets.every((t) => t.hcHealth === "healthy");
     const allUnhealthy = monitoredTargets.every(
         (t) => t.hcHealth === "unhealthy"
     );

server/lib/calculateUserClientsForOrgs.ts

@@ -339,19 +339,6 @@ export async function calculateUserClientsForOrgs(
                 continue;
             }
 
-            // Get exit nodes for this org
-            const exitNodesList = await getExitNodes(orgId);
-
-            if (exitNodesList.length === 0) {
-                logger.warn(
-                    `Skipping org ${orgId} for OLM ${olm.olmId} (user ${userId}): no exit nodes found`
-                );
-                continue;
-            }
-
-            const randomExitNode =
-                exitNodesList[Math.floor(Math.random() * exitNodesList.length)];
-
             // Get next available subnet
             const { value: newSubnet, release: releaseSubnetLock } =
                 await getNextAvailableClientSubnet(orgId, trx);
@@ -370,7 +357,6 @@ export async function calculateUserClientsForOrgs(
             const newClientData: InferInsertModel<typeof clients> = {
                 userId,
                 orgId: userOrg.orgId,
-                exitNodeId: randomExitNode.exitNodeId,
                 name: olm.name || "User Client",
                 subnet: updatedSubnet,
                 olmId: olm.olmId,

server/lib/deleteOrg.ts

@@ -93,6 +93,9 @@ export async function deleteOrgById(
             await trx.delete(sites).where(eq(sites.siteId, site.siteId));
         }
         for (const client of orgClients) {
+            if (client.exitNodeId && client.pubKey) {
+                await deletePeer(client.exitNodeId, client.pubKey);
+            }
             const [olm] = await trx
                 .select()
                 .from(olms)

server/lib/deleteResource.ts

@@ -64,13 +64,20 @@ export async function performDeleteResources(
 
     const targetsByResourceId = new Map<number, Target[]>();
     for (const target of targetsToBeRemoved) {
+        if (target.resourceId == null) {
+            continue;
+        }
         const existing = targetsByResourceId.get(target.resourceId) ?? [];
         existing.push(target);
         targetsByResourceId.set(target.resourceId, existing);
     }
 
     const targetIdToResourceId = new Map(
-        targetsToBeRemoved.map((target) => [target.targetId, target.resourceId])
+        targetsToBeRemoved.flatMap((target) =>
+            target.resourceId == null
+                ? []
+                : [[target.targetId, target.resourceId] as const]
+        )
     );
 
     const healthChecksByResourceId = new Map<number, TargetHealthCheck[]>();

server/lib/deleteSiteAssociatedResources.ts

@@ -1,4 +1,4 @@
-import { and, eq, inArray, sql } from "drizzle-orm";
+import { and, eq, inArray, isNotNull, sql } from "drizzle-orm";
 import {
     db,
     resources,
@@ -33,9 +33,11 @@ export async function getResourceIdsForSite(
     const rows = await trx
         .selectDistinct({ resourceId: targets.resourceId })
         .from(targets)
-        .where(eq(targets.siteId, siteId));
+        .where(and(eq(targets.siteId, siteId), isNotNull(targets.resourceId)));
 
-    return rows.map((row) => row.resourceId);
+    return rows
+        .map((row) => row.resourceId)
+        .filter((resourceId): resourceId is number => resourceId != null);
 }
 
 export async function getSiteResourceIdsForSite(

server/lib/exitNodes/calculateExitNodeWeight.ts

@@ -0,0 +1,40 @@
+import { db, sites, clients } from "@server/db";
+import { and, eq, count } from "drizzle-orm";
+
+// (MAX_CONNECTIONS - current_connections) / MAX_CONNECTIONS)
+// higher = more desirable
+// like saying, this node has x% of its capacity left
+export async function calculateExitNodeWeight(
+    exitNodeId: number,
+    maxConnections: number | null | undefined
+): Promise<number | null> {
+    if (maxConnections === null || maxConnections === undefined) {
+        return 1;
+    }
+
+    const [[siteConnections], [clientConnections]] = await Promise.all([
+        db
+            .select({ count: count() })
+            .from(sites)
+            .where(
+                and(eq(sites.exitNodeId, exitNodeId), eq(sites.online, true))
+            ),
+        db
+            .select({ count: count() })
+            .from(clients)
+            .where(
+                and(
+                    eq(clients.exitNodeId, exitNodeId),
+                    eq(clients.online, true)
+                )
+            )
+    ]);
+
+    const currentConnections = siteConnections.count + clientConnections.count;
+
+    if (currentConnections >= maxConnections) {
+        return null;
+    }
+
+    return (maxConnections - currentConnections) / maxConnections;
+}

server/lib/exitNodes/exitNodes.ts

@@ -1,6 +1,5 @@
 import { db, exitNodes, Transaction } from "@server/db";
 import logger from "@server/logger";
-import { ExitNodePingResult } from "@server/routers/newt";
 import { eq } from "drizzle-orm";
 
 export async function verifyExitNodeOrgAccess(
@@ -52,6 +51,16 @@ export async function listExitNodes(
     return allExitNodes;
 }
 
+export type ExitNodePingResult = {
+    exitNodeId: number;
+    latencyMs: number;
+    weight: number;
+    error?: string;
+    exitNodeName: string;
+    endpoint: string;
+    wasPreviouslyConnected: boolean;
+};
+
 export function selectBestExitNode(
     pingResults: ExitNodePingResult[]
 ): ExitNodePingResult | null {

server/lib/exitNodes/getUniqueSubnetForExitNode.ts

@@ -0,0 +1,41 @@
+import { db, ExitNode, Transaction, sites, clients } from "@server/db";
+import { eq } from "drizzle-orm";
+import config from "@server/lib/config";
+import { findNextAvailableCidr } from "@server/lib/ip";
+import { lockManager } from "#dynamic/lib/lock";
+
+export async function getUniqueSubnetForExitNode(
+    exitNode: ExitNode,
+    trx: Transaction | typeof db = db
+): Promise<string | null> {
+    const lockKey = `subnet-allocation:${exitNode.exitNodeId}`;
+
+    return await lockManager.withLock(
+        lockKey,
+        async () => {
+            const [sitesQuery, clientsQuery] = await Promise.all([
+                trx
+                    .select({ subnet: sites.exitNodeSubnet })
+                    .from(sites)
+                    .where(eq(sites.exitNodeId, exitNode.exitNodeId)),
+                trx
+                    .select({ subnet: clients.exitNodeSubnet })
+                    .from(clients)
+                    .where(eq(clients.exitNodeId, exitNode.exitNodeId))
+            ]);
+
+            const blockSize = config.getRawConfig().gerbil.site_block_size;
+            const subnets = [...sitesQuery, ...clientsQuery]
+                .map((row) => row.subnet)
+                .filter(
+                    (subnet): subnet is string =>
+                        !!subnet &&
+                        /^(\d{1,3}\.){3}\d{1,3}\/\d{1,2}$/.test(subnet)
+                );
+            subnets.push(exitNode.address.replace(/\/\d+$/, `/${blockSize}`));
+
+            return findNextAvailableCidr(subnets, blockSize, exitNode.address);
+        },
+        5000 // 5 second lock TTL - subnet allocation should be quick
+    );
+}

server/lib/exitNodes/index.ts

@@ -2,3 +2,5 @@ export * from "./exitNodes";
 export * from "./exitNodeComms";
 export * from "./subnet";
 export * from "./getCurrentExitNodeId";
+export * from "./calculateExitNodeWeight";
+export * from "./getUniqueSubnetForExitNode";

server/lib/ip.ts

@@ -528,7 +528,10 @@ export function generateRemoteSubnets(
 
 export type Alias = { alias: string | null; aliasAddress: string | null };
 
-export function generateAliasConfig(allSiteResources: SiteResource[]): Alias[] {
+export function generateAliasConfig(
+    allSiteResources: SiteResource[],
+    overrideIp?: string
+): Alias[] {
     return allSiteResources
         .filter(
             (sr) =>
@@ -539,7 +542,7 @@ export function generateAliasConfig(allSiteResources: SiteResource[]): Alias[] {
         )
         .map((sr) => ({
             alias: sr.alias || sr.fullDomain,
-            aliasAddress: sr.aliasAddress
+            aliasAddress: overrideIp || sr.aliasAddress
         }));
 }
 
@@ -660,9 +663,10 @@ export type CertRef = { id: string; cert: string; key: string };
  * certificate (e.g. a wildcard cert used by thousands of site resources)
  * only need that certificate sent once per sync message.
  */
-export function dedupeCertsForTargets(
-    targetsV2: SubnetProxyTargetV2[]
-): { targets: SubnetProxyTargetV2[]; certs: CertRef[] } {
+export function dedupeCertsForTargets(targetsV2: SubnetProxyTargetV2[]): {
+    targets: SubnetProxyTargetV2[];
+    certs: CertRef[];
+} {
     const idByContent = new Map<string, string>();
     const certs: CertRef[] = [];
 
@@ -674,7 +678,10 @@ export function dedupeCertsForTargets(
         const contentKey = `${target.tlsCert}|${target.tlsKey}`;
         let id = idByContent.get(contentKey);
         if (!id) {
-            id = createHash("sha1").update(contentKey).digest("hex").slice(0, 16);
+            id = createHash("sha1")
+                .update(contentKey)
+                .digest("hex")
+                .slice(0, 16);
             idByContent.set(contentKey, id);
             certs.push({ id, cert: target.tlsCert, key: target.tlsKey });
         }
@@ -708,7 +715,9 @@ export async function batchFetchCertsForSiteResources(
 ): Promise<CertByDomain> {
     const domains = new Set(
         allSiteResources
-            .filter((r) => r.enabled && r.mode === "http" && r.ssl && r.fullDomain)
+            .filter(
+                (r) => r.enabled && r.mode === "http" && r.ssl && r.fullDomain
+            )
             .map((r) => r.fullDomain as string)
     );
 
@@ -852,7 +861,11 @@ export async function generateSubnetProxyTargetV2(
                         new Set([siteResource.fullDomain]),
                         true
                     );
-                    if (certs.length > 0 && certs[0].certFile && certs[0].keyFile) {
+                    if (
+                        certs.length > 0 &&
+                        certs[0].certFile &&
+                        certs[0].keyFile
+                    ) {
                         tlsCert = certs[0].certFile;
                         tlsKey = certs[0].keyFile;
                     } else {

server/lib/readConfigFile.ts

@@ -79,7 +79,13 @@ export const configSchema = z
                     .default(3001)
                     .transform(stoi)
                     .pipe(portSchema),
+                ai_gateway_port: portSchema
+                    .optional()
+                    .default(3005)
+                    .transform(stoi)
+                    .pipe(portSchema),
                 badger_override: z.string().optional(),
+                ai_gateway_override: z.string().optional(),
                 next_port: portSchema
                     .optional()
                     .default(3002)
@@ -139,6 +145,7 @@ export const configSchema = z
                 integration_port: 3003,
                 external_port: 3000,
                 internal_port: 3001,
+                ai_gateway_port: 3005,
                 next_port: 3002,
                 internal_hostname: "pangolin",
                 session_cookie_name: "p_session_token",
@@ -394,6 +401,36 @@ export const configSchema = z
                 disable_enterprise_features: z.boolean().optional()
             })
             .optional(),
+        ai: z
+            .object({
+                model_catalog: z
+                    .object({
+                        upstream_url: z
+                            .url()
+                            .optional()
+                            .default("https://api.fossorial.io/api/v1/models"),
+                        // No default - only used when an operator wants to
+                        // pin the catalog to a local file instead of
+                        // fetching it from upstream_url.
+                        file: z.string().optional(),
+                        refresh_interval_min_hours: z
+                            .number()
+                            .positive()
+                            .gt(0)
+                            .optional()
+                            .default(6),
+                        refresh_interval_max_hours: z
+                            .number()
+                            .positive()
+                            .gt(0)
+                            .optional()
+                            .default(12)
+                    })
+                    .optional()
+                    .prefault({})
+            })
+            .optional()
+            .prefault({}),
         dns: z
             .object({
                 nameservers: z

server/lib/rebuildClientAssociations.ts

@@ -19,7 +19,7 @@ import {
     userOrgRoles,
     userSiteResources
 } from "@server/db";
-import { and, count, eq, inArray, ne } from "drizzle-orm";
+import { and, count, eq, inArray, isNotNull, ne } from "drizzle-orm";
 
 import { deletePeersBatch as newtDeletePeersBatch } from "@server/routers/newt/peers";
 import {
@@ -27,6 +27,9 @@ import {
     deletePeersBatch as olmDeletePeersBatch
 } from "@server/routers/olm/peers";
 import { sendToExitNode } from "#dynamic/lib/exitNodes";
+import { sendToClientsBatch } from "#dynamic/routers/ws";
+import { canCompress } from "@server/lib/clientVersionChecks";
+import config from "@server/lib/config";
 import logger from "@server/logger";
 import {
     generateAliasConfig,
@@ -187,7 +190,12 @@ export async function getClientSiteResourceAccess(
         `rebuildClientAssociations: [getClientSiteResourceAccess] siteResourceId=${siteResource.siteResourceId} networkId=${siteResource.networkId} siteCount=${sitesList.length} siteIds=[${sitesList.map((s) => s.siteId).join(", ")}]`
     );
 
-    if (sitesList.length === 0) {
+    if (sitesList.length === 0 && siteResource.networkId !== null) {
+        // A site resource with a networkId is expected to have at least one
+        // site attached via siteNetworks. Resources with no networkId (e.g.
+        // inference-mode resources, which connect clients directly to the
+        // exit node instead of any site) are expected to have no sites, so
+        // don't warn for those.
         logger.warn(
             `No sites found for siteResource ${siteResource.siteResourceId} with networkId ${siteResource.networkId}`
         );
@@ -687,6 +695,22 @@ async function rebuildClientAssociationsFromSiteResourceImpl(
         clientSiteResourcesToRemove,
         trx
     );
+
+    // If this resource requires clients to be connected to the exit node
+    // (e.g. an inference resource), re-sync the connect/disconnect state for
+    // every client whose access to it may have changed - both those who
+    // currently have access and those who just lost it.
+    if (siteResource.requiresExitNodeConnection) {
+        await syncClientExitNodeConnections(
+            Array.from(
+                new Set([
+                    ...mergedAllClientIds,
+                    ...existingClientSiteResourceIds
+                ])
+            ),
+            trx
+        );
+    }
 }
 
 async function handleMessagesForSiteClients(
@@ -966,7 +990,7 @@ export async function updateClientSiteDestinations(
         .where(eq(clientSitesAssociationsCache.clientId, client.clientId));
 
     for (const site of sitesData) {
-        if (!site.sites.subnet) {
+        if (!site.sites.exitNodeSubnet) {
             logger.debug(`Site ${site.sites.siteId} has no subnet, skipping`);
             continue;
         }
@@ -1002,7 +1026,7 @@ export async function updateClientSiteDestinations(
                 sourcePort: parsedEndpoint.port,
                 destinations: [
                     {
-                        destinationIP: site.sites.subnet.split("/")[0],
+                        destinationIP: site.sites.exitNodeSubnet.split("/")[0],
                         destinationPort: site.sites.listenPort || 1 // this satisfies gerbil for now but should be reevaluated
                     }
                 ]
@@ -1010,7 +1034,7 @@ export async function updateClientSiteDestinations(
         } else {
             // add to the existing destinations
             destinations.destinations.push({
-                destinationIP: site.sites.subnet.split("/")[0],
+                destinationIP: site.sites.exitNodeSubnet.split("/")[0],
                 destinationPort: site.sites.listenPort || 1 // this satisfies gerbil for now but should be reevaluated
             });
         }
@@ -1052,6 +1076,265 @@ export async function updateClientSiteDestinations(
     }
 }
 
+// Determines, for each of the given clients, whether they currently have
+// access to any enabled site resource with requiresExitNodeConnection set
+// (e.g. an inference-mode resource) and tells the client's olm to connect to
+// or disconnect from its assigned exit node accordingly. Site resources with
+// requiresExitNodeConnection don't belong to any site/network, so this can't
+// be derived from the per-site peer logic above - it has to be recomputed
+// from the client's full current resource access every time that access
+// changes.
+async function syncClientExitNodeConnections(
+    clientIds: number[],
+    trx: Transaction | typeof db = db
+): Promise<void> {
+    const uniqueClientIds = Array.from(new Set(clientIds));
+    if (uniqueClientIds.length === 0) {
+        return;
+    }
+
+    // Only clients with an exit node assigned can be told to connect/disconnect.
+    const clientsData = await trx
+        .select({
+            clientId: clients.clientId,
+            exitNodeId: clients.exitNodeId,
+            exitNodeSubnet: clients.exitNodeSubnet
+        })
+        .from(clients)
+        .where(
+            and(
+                inArray(clients.clientId, uniqueClientIds),
+                isNotNull(clients.exitNodeId)
+            )
+        );
+
+    if (clientsData.length === 0) {
+        return;
+    }
+
+    const clientIdsWithExitNode = clientsData.map((c) => c.clientId);
+
+    const requiresExitNodeRows = await trx
+        .select({
+            clientId: clientSiteResourcesAssociationsCache.clientId,
+            alias: siteResources.alias,
+            fullDomain: siteResources.fullDomain
+        })
+        .from(clientSiteResourcesAssociationsCache)
+        .innerJoin(
+            siteResources,
+            eq(
+                clientSiteResourcesAssociationsCache.siteResourceId,
+                siteResources.siteResourceId
+            )
+        )
+        .where(
+            and(
+                inArray(
+                    clientSiteResourcesAssociationsCache.clientId,
+                    clientIdsWithExitNode
+                ),
+                eq(siteResources.enabled, true),
+                eq(siteResources.requiresExitNodeConnection, true)
+            )
+        );
+
+    const needsConnectSet = new Set(
+        requiresExitNodeRows.map((r) => r.clientId)
+    );
+
+    // Aliases for every exit-node-backed resource this client can reach, so
+    // the live connect push carries the same alias list the register/reconnect
+    // path (buildSiteConfigurationForOlmClient) would compute.
+    const exitNodeAliasesByClientId = new Map<number, (string | null)[]>();
+    for (const row of requiresExitNodeRows) {
+        if (row.alias == null && row.fullDomain == null) continue;
+        const existing = exitNodeAliasesByClientId.get(row.clientId);
+        if (existing) {
+            existing.push(row.fullDomain || row.alias); // accept both for now in case we have other resource types that dont use the full domain
+        } else {
+            exitNodeAliasesByClientId.set(row.clientId, [
+                row.fullDomain || row.alias
+            ]);
+        }
+    }
+
+    const exitNodeIds = Array.from(
+        new Set(
+            clientsData
+                .map((c) => c.exitNodeId)
+                .filter((id): id is number => id !== null)
+        )
+    );
+
+    const exitNodeRows =
+        exitNodeIds.length > 0
+            ? await trx
+                  .select()
+                  .from(exitNodes)
+                  .where(inArray(exitNodes.exitNodeId, exitNodeIds))
+            : [];
+    const exitNodeById = new Map(exitNodeRows.map((n) => [n.exitNodeId, n]));
+
+    const olmRows = await trx
+        .select({
+            clientId: olms.clientId,
+            olmId: olms.olmId,
+            version: olms.version
+        })
+        .from(olms)
+        .where(inArray(olms.clientId, clientIdsWithExitNode));
+    const olmByClientId = new Map(
+        olmRows
+            .filter((r) => r.clientId !== null)
+            .map((r) => [r.clientId as number, r])
+    );
+
+    const relayPort = config.getRawConfig().gerbil.clients_start_port;
+
+    const connectPayloads: {
+        clientId: string;
+        message: { type: string; data: any };
+        options: { compress: boolean; incrementConfigVersion: boolean };
+    }[] = [];
+    const disconnectPayloads: {
+        clientId: string;
+        message: { type: string; data: any };
+        options: { compress: boolean; incrementConfigVersion: boolean };
+    }[] = [];
+
+    for (const client of clientsData) {
+        const olm = olmByClientId.get(client.clientId);
+        if (!olm) {
+            // No olm registered for this client yet/anymore, nothing to send.
+            continue;
+        }
+
+        const needsConnect = needsConnectSet.has(client.clientId);
+
+        if (needsConnect) {
+            const exitNode = client.exitNodeId
+                ? exitNodeById.get(client.exitNodeId)
+                : undefined;
+            if (!exitNode || !client.exitNodeSubnet) {
+                logger.warn(
+                    `rebuildClientAssociations: [syncClientExitNodeConnections] client ${client.clientId} needs an exit node connection but has no exit node or subnet assigned`
+                );
+                continue;
+            }
+
+            connectPayloads.push({
+                clientId: olm.olmId,
+                message: {
+                    type: "olm/wg/exitnode/connect",
+                    data: {
+                        connect: true,
+                        endpoint: `${exitNode.endpoint}:${exitNode.listenPort}`,
+                        relayPort,
+                        publicKey: exitNode.publicKey,
+                        serverIP: exitNode.address.split("/")[0],
+                        tunnelIP: client.exitNodeSubnet.split("/")[0],
+                        aliases:
+                            exitNodeAliasesByClientId.get(client.clientId) ?? []
+                    }
+                },
+                options: {
+                    compress: canCompress(olm.version, "olm"),
+                    incrementConfigVersion: true
+                }
+            });
+        } else {
+            disconnectPayloads.push({
+                clientId: olm.olmId,
+                message: {
+                    type: "olm/wg/exitnode/disconnect",
+                    data: {}
+                },
+                options: {
+                    compress: canCompress(olm.version, "olm"),
+                    incrementConfigVersion: true
+                }
+            });
+        }
+    }
+
+    if (connectPayloads.length > 0) {
+        await sendToClientsBatch(connectPayloads).catch((error) => {
+            logger.error(
+                `rebuildClientAssociations: Error sending exit node connect messages:`,
+                error
+            );
+        });
+    }
+
+    if (disconnectPayloads.length > 0) {
+        await sendToClientsBatch(disconnectPayloads).catch((error) => {
+            logger.error(
+                `rebuildClientAssociations: Error sending exit node disconnect messages:`,
+                error
+            );
+        });
+    }
+}
+
+// Notifies the olms of every given client that the alias of the site resource
+// they're using an exit node connection for has changed, via the dedicated
+// exit node data-update message. Unlike syncClientExitNodeConnections, this
+// doesn't touch connect/disconnect state - it's purely a rename for clients
+// that are (and remain) connected to the exit node for this resource.
+async function syncClientExitNodeAliasUpdate(
+    clientIds: number[],
+    oldAlias: string | null,
+    newAlias: string | null,
+    trx: Transaction | typeof db = db
+): Promise<void> {
+    const uniqueClientIds = Array.from(new Set(clientIds));
+    if (uniqueClientIds.length === 0) {
+        return;
+    }
+
+    const oldAliases = oldAlias ? [oldAlias] : [];
+    const newAliases = newAlias ? [newAlias] : [];
+    if (oldAliases.length === 0 && newAliases.length === 0) {
+        return;
+    }
+
+    const olmRows = await trx
+        .select({
+            clientId: olms.clientId,
+            olmId: olms.olmId,
+            version: olms.version
+        })
+        .from(olms)
+        .where(inArray(olms.clientId, uniqueClientIds));
+
+    const updatePayloads = olmRows
+        .filter((r) => r.clientId !== null)
+        .map((olm) => ({
+            clientId: olm.olmId,
+            message: {
+                type: "olm/wg/exitnode/data/update",
+                data: {
+                    oldAliases,
+                    newAliases
+                }
+            },
+            options: {
+                compress: canCompress(olm.version, "olm"),
+                incrementConfigVersion: true // this is important information we would need to sync
+            }
+        }));
+
+    if (updatePayloads.length > 0) {
+        await sendToClientsBatch(updatePayloads).catch((error) => {
+            logger.error(
+                `rebuildClientAssociations: Error sending exit node alias update messages:`,
+                error
+            );
+        });
+    }
+}
+
 async function handleSubnetProxyTargetUpdates(
     siteResource: SiteResource,
     sitesList: Site[],
@@ -1282,7 +1565,7 @@ export async function handleMessagingForUpdatedSiteResource(
         `handleMessagingForUpdatedSiteResource: fetched newts for ${newtsForSites.length}/${allSiteIds.length} site(s)`
     );
 
-    // WARNING: THIS RELIES ON THE CACHE TABLES BEING UP TO DATE, SO CALL THIS AFTER THE ASSOCIATION CACHE IS UPDATED
+    // !!!!!!!!!!!!!!!!!! WARNING: THIS RELIES ON THE CACHE TABLES BEING UP TO DATE, SO CALL THIS AFTER THE ASSOCIATION CACHE IS UPDATED !!!!!!!!!!!!!!!!!!
     const mergedAllClients = await trx
         .select({
             clientId: clientSiteResourcesAssociationsCache.clientId,
@@ -1709,6 +1992,38 @@ export async function handleMessagingForUpdatedSiteResource(
         );
     }
 
+    // For a resource that stays on an exit node connection across the update,
+    // the alias is the only field that affects already-connected clients (the
+    // exit node itself, its endpoint, etc. are not per-resource). Tell those
+    // clients' olms about the rename directly via the exit node data-update
+    // message rather than a full connect/disconnect cycle.
+    if (
+        existingSiteResource?.requiresExitNodeConnection &&
+        updatedSiteResource.requiresExitNodeConnection &&
+        aliasChanged
+    ) {
+        await syncClientExitNodeAliasUpdate(
+            mergedAllClients.map((c) => c.clientId),
+            existingSiteResource.alias,
+            updatedSiteResource.alias,
+            trx
+        );
+    }
+
+    // If this resource requires (or required) clients to be connected to the
+    // exit node (e.g. an inference resource), re-sync connect/disconnect
+    // state for every client currently associated with it - covers toggling
+    // requiresExitNodeConnection on update as well as enabling/disabling it.
+    if (
+        updatedSiteResource.requiresExitNodeConnection ||
+        existingSiteResource?.requiresExitNodeConnection
+    ) {
+        await syncClientExitNodeConnections(
+            mergedAllClients.map((c) => c.clientId),
+            trx
+        );
+    }
+
     logger.debug(
         `handleMessagingForUpdatedSiteResource: DONE siteResourceId=${updatedSiteResource.siteResourceId}`
     );
@@ -1990,6 +2305,10 @@ async function rebuildClientAssociationsFromClientImpl(
         resourcesToRemove,
         trx
     );
+
+    // Re-sync exit node connect/disconnect state based on this client's
+    // current full set of resource access (e.g. inference resources).
+    await syncClientExitNodeConnections([client.clientId], trx);
 }
 
 async function handleMessagesForClientSites(

server/lib/traefik/TraefikConfigManager.ts

@@ -516,6 +516,11 @@ export class TraefikConfigManager {
             const maintenanceHost =
                 config.getRawConfig().server.internal_hostname;
             const pangolinUIUrl = `http://${maintenanceHost}:${maintenancePort}`;
+            const aiGatewayUrl =
+                config.getRawConfig().server.ai_gateway_override ||
+                `http://${maintenanceHost}:${
+                    config.getRawConfig().server.ai_gateway_port
+                }`;
 
             // logger.debug(`Fetching traefik config for exit node: ${currentExitNode}`);
             traefikConfig = await getTraefikConfig(
@@ -528,7 +533,8 @@ export class TraefikConfigManager {
                     ? false
                     : config.getRawConfig().traefik.allow_raw_resources, // dont allow raw resources on saas otherwise use config
                 pangolinUIUrl, // generate maintenance pages on cloud and hybrid
-                pangolinUIUrl // generate browser gateway targets on cloud and hybrid
+                pangolinUIUrl, // generate browser gateway targets on cloud and hybrid
+                aiGatewayUrl
             );
 
             const domains = new Set<string>();

server/lib/traefik/getTraefikConfig.ts

@@ -1,4 +1,12 @@
-import { db, targetHealthCheck, domains } from "@server/db";
+import {
+    db,
+    targetHealthCheck,
+    domains,
+    aiProviders,
+    resourceAiProviders,
+    siteResources,
+    exitNodes
+} from "@server/db";
 import {
     and,
     eq,
@@ -15,6 +23,7 @@ import config from "@server/lib/config";
 import { resources, sites, Target, targets } from "@server/db";
 import createPathRewriteMiddleware from "./middleware";
 import { sanitize, encodePath, validatePathRewriteConfig } from "./utils";
+import regionalCache from "@server/lib/cache";
 
 const redirectHttpsMiddlewareName = "redirect-to-https";
 const badgerMiddlewareName = "badger";
@@ -45,8 +54,24 @@ export async function getTraefikConfig(
     generateLoginPageRouters = false, // UNUSED BUT USED IN PRIVATE
     allowRawResources = true,
     maintenancePageUiUrl: string | null = null, // UNUSED BUT USED IN PRIVATE
-    browserGatewayUiUrl: string | null = null // UNUSED BUT USED IN PRIVATE
+    browserGatewayUiUrl: string | null = null, // UNUSED BUT USED IN PRIVATE
+    aiGatewayUrl: string | null = null
 ): Promise<any> {
+    // Get the exit node but cache it for 5 minutes to avoid hitting the DB too often
+    const exitNodeCacheKey = `exitNode:${exitNodeId}`;
+    let exitNode =
+        await regionalCache.get<typeof exitNodes.$inferSelect>(
+            exitNodeCacheKey
+        );
+    if (!exitNode) {
+        [exitNode] = await db
+            .select()
+            .from(exitNodes)
+            .where(eq(exitNodes.exitNodeId, exitNodeId))
+            .limit(1);
+        await regionalCache.set(exitNodeCacheKey, exitNode, 300);
+    }
+
     // Get resources with their targets and sites in a single optimized query
     // Start from sites on this exit node, then join to targets and resources
     const resourcesWithTargetsAndSites = await db
@@ -87,7 +112,7 @@ export async function getTraefikConfig(
             siteId: sites.siteId,
             siteType: sites.type,
             siteOnline: sites.online,
-            subnet: sites.subnet,
+            subnet: sites.exitNodeSubnet,
             exitNodeId: sites.exitNodeId,
             // Domain cert resolver fields
             domainCertResolver: domains.certResolver,
@@ -209,8 +234,42 @@ export async function getTraefikConfig(
         });
     });
 
+    // Inference-mode resources have no targets/sites (their "backend" is the
+    // central AI gateway), so they can't be reached via the targets->sites
+    // join above - query them separately and include them on every exit node.
+    const inferenceResources = await db
+        .selectDistinct({
+            resourceId: resources.resourceId,
+            resourceName: resources.name,
+            fullDomain: resources.fullDomain,
+            ssl: resources.ssl,
+            subdomain: resources.subdomain,
+            domainId: resources.domainId,
+            enabled: resources.enabled,
+            wildcard: resources.wildcard,
+            domainCertResolver: domains.certResolver,
+            preferWildcardCert: domains.preferWildcardCert
+        })
+        .from(resources)
+        // .innerJoin(
+        //     resourceAiProviders,
+        //     eq(resources.resourceId, resourceAiProviders.resourceId)
+        // )
+        // .innerJoin(
+        //     aiProviders,
+        //     eq(resourceAiProviders.providerId, aiProviders.providerId)
+        // )
+        .leftJoin(domains, eq(domains.domainId, resources.domainId))
+        .where(
+            and(
+                eq(resources.mode, "inference"),
+                eq(resources.enabled, true)
+                // eq(aiProviders.enabled, true)
+            )
+        );
+
     // make sure we have at least one resource
-    if (resourcesMap.size === 0) {
+    if (resourcesMap.size === 0 && inferenceResources.length === 0) {
         return {};
     }
 
@@ -673,5 +732,229 @@ export async function getTraefikConfig(
             };
         }
     }
+
+    if (aiGatewayUrl) {
+        // The AI gateway may live on a different host than the inference
+        // resource itself (e.g. a remote exit node forwarding to the
+        // central dashboard over a tunnel). passHostHeader would forward
+        // the resource's own Host, which that external host won't
+        // recognize, so we pin the Host header to the gateway's own host
+        // and smuggle the original resource host through in "p-host"
+        // instead.
+        let aiGatewayHost: string | undefined;
+        try {
+            aiGatewayHost = new URL(aiGatewayUrl).host;
+        } catch {
+            aiGatewayHost = undefined;
+        }
+
+        // Public inference resources: same TLS/cert-resolver handling as
+        // plain http-mode resources, but the service points at the AI
+        // gateway instead of any real backend targets.
+        for (const ir of inferenceResources) {
+            if (!ir.enabled) continue;
+            if (!ir.domainId || !ir.fullDomain) continue;
+
+            if (!config_output.http.routers) config_output.http.routers = {};
+            if (!config_output.http.services) config_output.http.services = {};
+
+            const fullDomain = ir.fullDomain;
+            const irKey = `inference-r${ir.resourceId}`;
+            const routerName = `${irKey}-router`;
+            const serviceName = `${irKey}-service`;
+
+            let rule: string;
+            if (ir.wildcard && fullDomain.startsWith("*.")) {
+                const escaped = fullDomain.slice(2).replace(/\./g, "\\.");
+                rule = `HostRegexp(\`^[^.]+\\.${escaped}$\`)`;
+            } else {
+                rule = `Host(\`${fullDomain}\`)`;
+            }
+
+            const domainParts = fullDomain.split(".");
+            let wildCard;
+            if (domainParts.length <= 2) {
+                wildCard = `*.${domainParts.join(".")}`;
+            } else {
+                wildCard = `*.${domainParts.slice(1).join(".")}`;
+            }
+            if (!ir.subdomain) {
+                wildCard = fullDomain;
+            }
+
+            const globalDefaultResolver =
+                config.getRawConfig().traefik.cert_resolver;
+            const globalDefaultPreferWildcard =
+                config.getRawConfig().traefik.prefer_wildcard_cert;
+            const resolverName = ir.domainCertResolver
+                ? ir.domainCertResolver.trim()
+                : globalDefaultResolver;
+            const preferWildcard =
+                ir.preferWildcardCert !== undefined &&
+                ir.preferWildcardCert !== null
+                    ? ir.preferWildcardCert
+                    : globalDefaultPreferWildcard;
+
+            const tls = {
+                certResolver: resolverName,
+                ...(preferWildcard ? { domains: [{ main: wildCard }] } : {})
+            };
+
+            const irHeadersMiddlewareName = `${irKey}-headers-middleware`;
+            if (!config_output.http.middlewares) {
+                config_output.http.middlewares = {};
+            }
+            config_output.http.middlewares[irHeadersMiddlewareName] = {
+                headers: {
+                    customRequestHeaders: {
+                        ...(aiGatewayHost ? { Host: aiGatewayHost } : {}),
+                        "p-host": fullDomain
+                    }
+                }
+            };
+
+            const additionalMiddlewares =
+                config.getRawConfig().traefik.additional_middlewares || [];
+            const routerMiddlewares = [
+                badgerMiddlewareName,
+                irHeadersMiddlewareName,
+                ...additionalMiddlewares
+            ];
+
+            if (ir.ssl) {
+                config_output.http.routers[routerName + "-redirect"] = {
+                    entryPoints: [
+                        config.getRawConfig().traefik.http_entrypoint
+                    ],
+                    middlewares: [redirectHttpsMiddlewareName],
+                    service: serviceName,
+                    rule,
+                    priority: 100
+                };
+            }
+
+            config_output.http.routers[routerName] = {
+                entryPoints: [
+                    ir.ssl
+                        ? config.getRawConfig().traefik.https_entrypoint
+                        : config.getRawConfig().traefik.http_entrypoint
+                ],
+                middlewares: routerMiddlewares,
+                service: serviceName,
+                rule,
+                priority: 100,
+                ...(ir.ssl ? { tls } : {})
+            };
+
+            config_output.http.services[serviceName] = {
+                loadBalancer: {
+                    servers: [{ url: aiGatewayUrl }]
+                }
+            };
+        }
+
+        // Private (siteResource) inference resources: routed by their alias
+        // instead of a public fullDomain, and deliberately WITHOUT the
+        // badger middleware - no per-user auth/policy stack exists for
+        // siteResources today, so gating here is reachability-only for now.
+        const siteResourcesInference = await db
+            .selectDistinct({
+                siteResourceId: siteResources.siteResourceId,
+                fullDomain: siteResources.fullDomain,
+                ssl: siteResources.ssl,
+                enabled: siteResources.enabled
+            })
+            .from(siteResources)
+            .where(
+                and(
+                    eq(siteResources.mode, "inference"),
+                    eq(siteResources.enabled, true),
+                    isNotNull(siteResources.fullDomain)
+                )
+            );
+
+        for (const sr of siteResourcesInference) {
+            if (!sr.enabled || !sr.fullDomain) continue;
+
+            if (!config_output.http.routers) config_output.http.routers = {};
+            if (!config_output.http.services) config_output.http.services = {};
+
+            const fullDomain = sr.fullDomain;
+            const srKey = `inference-sr${sr.siteResourceId}`;
+            const routerName = `${srKey}-router`;
+            const serviceName = `${srKey}-service`;
+            const rule = `Host(\`${fullDomain}\`) && ClientIP(${exitNode.address})`; // restrict to coming from the exit node ip range that the client is connected to
+
+            const domainParts = fullDomain.split(".");
+            const wildCard =
+                domainParts.length <= 2
+                    ? `*.${domainParts.join(".")}`
+                    : `*.${domainParts.slice(1).join(".")}`;
+
+            const globalDefaultResolver =
+                config.getRawConfig().traefik.cert_resolver;
+            const globalDefaultPreferWildcard =
+                config.getRawConfig().traefik.prefer_wildcard_cert;
+
+            const tls = {
+                certResolver: globalDefaultResolver,
+                ...(globalDefaultPreferWildcard
+                    ? { domains: [{ main: wildCard }] }
+                    : {})
+            };
+
+            const srHeadersMiddlewareName = `${srKey}-headers-middleware`;
+            if (!config_output.http.middlewares) {
+                config_output.http.middlewares = {};
+            }
+            config_output.http.middlewares[srHeadersMiddlewareName] = {
+                headers: {
+                    customRequestHeaders: {
+                        ...(aiGatewayHost ? { Host: aiGatewayHost } : {}),
+                        "p-host": fullDomain
+                    }
+                }
+            };
+
+            const additionalMiddlewares =
+                config.getRawConfig().traefik.additional_middlewares || [];
+            const routerMiddlewares = [
+                srHeadersMiddlewareName,
+                ...additionalMiddlewares
+            ];
+
+            if (sr.ssl) {
+                config_output.http.routers[routerName + "-redirect"] = {
+                    entryPoints: [
+                        config.getRawConfig().traefik.http_entrypoint
+                    ],
+                    middlewares: [redirectHttpsMiddlewareName],
+                    service: serviceName,
+                    rule,
+                    priority: 100
+                };
+            }
+
+            config_output.http.routers[routerName] = {
+                entryPoints: [
+                    sr.ssl
+                        ? config.getRawConfig().traefik.https_entrypoint
+                        : config.getRawConfig().traefik.http_entrypoint
+                ],
+                middlewares: routerMiddlewares,
+                service: serviceName,
+                rule,
+                priority: 100,
+                ...(sr.ssl ? { tls } : {})
+            };
+
+            config_output.http.services[serviceName] = {
+                loadBalancer: {
+                    servers: [{ url: aiGatewayUrl }]
+                }
+            };
+        }
+    }
+
     return config_output;
 }

server/middlewares/index.ts

@@ -27,6 +27,9 @@ export * from "./verifyUserHasAction";
 export * from "./verifyApiKeyAccess";
 export * from "./verifySiteProvisioningKeyAccess";
 export * from "./verifyDomainAccess";
+export * from "./verifyAiProviderAccess";
+export * from "./verifyAiModelAccess";
+export * from "./verifyAiBudgetAccess";
 export * from "./verifyUserIsOrgOwner";
 export * from "./verifyUserFromResourceSession";
 export * from "./verifySiteResourceAccess";

server/middlewares/integration/index.ts

@@ -16,5 +16,7 @@ export * from "./verifyApiKeyClientAccess";
 export * from "./verifyApiKeySiteResourceAccess";
 export * from "./verifyApiKeyIdpAccess";
 export * from "./verifyApiKeyDomainAccess";
+export * from "./verifyApiKeyAiProviderAccess";
+export * from "./verifyApiKeyAiModelAccess";
 export * from "./verifyApiKeyResourcePolicyAccess";
 export * from "./verifyApiKeySiteProvisioningKeyAccess";

server/middlewares/integration/verifyApiKeyAiModelAccess.ts

@@ -0,0 +1,94 @@
+import { Request, Response, NextFunction } from "express";
+import { aiModels, aiProviders, apiKeyOrg, db } from "@server/db";
+import { and, eq } from "drizzle-orm";
+import createHttpError from "http-errors";
+import HttpCode from "@server/types/HttpCode";
+import { getFirstString } from "@server/lib/requestParams";
+
+export async function verifyApiKeyAiModelAccess(
+    req: Request,
+    res: Response,
+    next: NextFunction
+) {
+    try {
+        const apiKey = req.apiKey;
+        const modelIdRaw = getFirstString(req.params.modelId);
+        const modelId = Number.parseInt(modelIdRaw ?? "", 10);
+
+        if (!apiKey) {
+            return next(
+                createHttpError(HttpCode.UNAUTHORIZED, "Key not authenticated")
+            );
+        }
+
+        if (Number.isNaN(modelId)) {
+            return next(
+                createHttpError(HttpCode.BAD_REQUEST, "Invalid model ID")
+            );
+        }
+
+        const [row] = await db
+            .select({
+                model: aiModels,
+                provider: aiProviders
+            })
+            .from(aiModels)
+            .innerJoin(
+                aiProviders,
+                eq(aiModels.providerId, aiProviders.providerId)
+            )
+            .where(eq(aiModels.modelId, modelId))
+            .limit(1);
+
+        if (!row) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI model with ID ${modelId} not found`
+                )
+            );
+        }
+
+        if (apiKey.isRoot) {
+            req.aiProvider = row.provider;
+            req.aiModel = row.model;
+            return next();
+        }
+
+        const orgId = row.provider.orgId;
+
+        if (!req.apiKeyOrg || req.apiKeyOrg.orgId !== orgId) {
+            const apiKeyOrgRes = await db
+                .select()
+                .from(apiKeyOrg)
+                .where(
+                    and(
+                        eq(apiKeyOrg.apiKeyId, apiKey.apiKeyId),
+                        eq(apiKeyOrg.orgId, orgId)
+                    )
+                )
+                .limit(1);
+            req.apiKeyOrg = apiKeyOrgRes[0];
+        }
+
+        if (!req.apiKeyOrg) {
+            return next(
+                createHttpError(
+                    HttpCode.FORBIDDEN,
+                    "Key does not have access to this organization"
+                )
+            );
+        }
+
+        req.aiProvider = row.provider;
+        req.aiModel = row.model;
+        return next();
+    } catch (error) {
+        return next(
+            createHttpError(
+                HttpCode.INTERNAL_SERVER_ERROR,
+                "Error verifying AI model access"
+            )
+        );
+    }
+}

server/middlewares/integration/verifyApiKeyAiProviderAccess.ts

@@ -0,0 +1,85 @@
+import { Request, Response, NextFunction } from "express";
+import { aiProviders, apiKeyOrg, db } from "@server/db";
+import { and, eq } from "drizzle-orm";
+import createHttpError from "http-errors";
+import HttpCode from "@server/types/HttpCode";
+import { getFirstString } from "@server/lib/requestParams";
+
+export async function verifyApiKeyAiProviderAccess(
+    req: Request,
+    res: Response,
+    next: NextFunction
+) {
+    try {
+        const apiKey = req.apiKey;
+        const providerIdRaw = getFirstString(req.params.providerId);
+        const providerId = Number.parseInt(providerIdRaw ?? "", 10);
+
+        if (!apiKey) {
+            return next(
+                createHttpError(HttpCode.UNAUTHORIZED, "Key not authenticated")
+            );
+        }
+
+        if (Number.isNaN(providerId)) {
+            return next(
+                createHttpError(HttpCode.BAD_REQUEST, "Invalid provider ID")
+            );
+        }
+
+        const [provider] = await db
+            .select()
+            .from(aiProviders)
+            .where(eq(aiProviders.providerId, providerId))
+            .limit(1);
+
+        if (!provider) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI provider with ID ${providerId} not found`
+                )
+            );
+        }
+
+        if (apiKey.isRoot) {
+            req.aiProvider = provider;
+            return next();
+        }
+
+        const orgId = provider.orgId;
+
+        if (!req.apiKeyOrg || req.apiKeyOrg.orgId !== orgId) {
+            const apiKeyOrgRes = await db
+                .select()
+                .from(apiKeyOrg)
+                .where(
+                    and(
+                        eq(apiKeyOrg.apiKeyId, apiKey.apiKeyId),
+                        eq(apiKeyOrg.orgId, orgId)
+                    )
+                )
+                .limit(1);
+            req.apiKeyOrg = apiKeyOrgRes[0];
+        }
+
+        if (!req.apiKeyOrg) {
+            return next(
+                createHttpError(
+                    HttpCode.FORBIDDEN,
+                    "Key does not have access to this organization"
+                )
+            );
+        }
+
+        req.aiProvider = provider;
+        return next();
+    } catch (error) {
+        return next(
+            createHttpError(
+                HttpCode.INTERNAL_SERVER_ERROR,
+                "Error verifying AI provider access"
+            )
+        );
+    }
+}

server/middlewares/integration/verifyApiKeyTargetAccess.ts

@@ -1,6 +1,6 @@
 import { Request, Response, NextFunction } from "express";
 import { db } from "@server/db";
-import { resources, targets, apiKeyOrg } from "@server/db";
+import { aiProviders, resources, targets, apiKeyOrg } from "@server/db";
 import { and, eq } from "drizzle-orm";
 import createHttpError from "http-errors";
 import HttpCode from "@server/types/HttpCode";
@@ -43,43 +43,65 @@ export async function verifyApiKeyTargetAccess(
             );
         }
 
-        const resourceId = target.resourceId;
-        if (!resourceId) {
+        const { resourceId, providerId } = target;
+        if ((!resourceId && !providerId) || (resourceId && providerId)) {
             return next(
                 createHttpError(
                     HttpCode.INTERNAL_SERVER_ERROR,
-                    `Target with ID ${targetId} does not have a resource ID`
-                )
-            );
-        }
-
-        const [resource] = await db
-            .select()
-            .from(resources)
-            .where(eq(resources.resourceId, resourceId))
-            .limit(1);
-
-        if (!resource) {
-            return next(
-                createHttpError(
-                    HttpCode.NOT_FOUND,
-                    `Resource with ID ${resourceId} not found`
+                    `Target with ID ${targetId} has invalid ownership`
                 )
             );
         }
 
         if (apiKey.isRoot) {
-            // Root keys can access any key in any org
+            // Root keys can access any target
             return next();
         }
 
-        if (!resource.orgId) {
-            return next(
-                createHttpError(
-                    HttpCode.INTERNAL_SERVER_ERROR,
-                    `Resource with ID ${resourceId} does not have an organization ID`
-                )
-            );
+        let orgId: string;
+        if (resourceId) {
+            const [resource] = await db
+                .select()
+                .from(resources)
+                .where(eq(resources.resourceId, resourceId))
+                .limit(1);
+
+            if (!resource) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `Resource with ID ${resourceId} not found`
+                    )
+                );
+            }
+
+            if (!resource.orgId) {
+                return next(
+                    createHttpError(
+                        HttpCode.INTERNAL_SERVER_ERROR,
+                        `Resource with ID ${resourceId} does not have an organization ID`
+                    )
+                );
+            }
+
+            orgId = resource.orgId;
+        } else {
+            const [provider] = await db
+                .select()
+                .from(aiProviders)
+                .where(eq(aiProviders.providerId, providerId!))
+                .limit(1);
+
+            if (!provider) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `AI provider with ID ${providerId} not found`
+                    )
+                );
+            }
+
+            orgId = provider.orgId;
         }
 
         if (!req.apiKeyOrg) {
@@ -89,7 +111,7 @@ export async function verifyApiKeyTargetAccess(
                 .where(
                     and(
                         eq(apiKeyOrg.apiKeyId, apiKey.apiKeyId),
-                        eq(apiKeyOrg.orgId, resource.orgId)
+                        eq(apiKeyOrg.orgId, orgId)
                     )
                 )
                 .limit(1);
@@ -98,7 +120,7 @@ export async function verifyApiKeyTargetAccess(
             }
         }
 
-        if (!req.apiKeyOrg) {
+        if (!req.apiKeyOrg || req.apiKeyOrg.orgId !== orgId) {
             return next(
                 createHttpError(
                     HttpCode.FORBIDDEN,

server/middlewares/verifyAiBudgetAccess.ts

@@ -0,0 +1,99 @@
+import { Request, Response, NextFunction } from "express";
+import { aiBudgets, db, userOrgs } from "@server/db";
+import { and, eq } from "drizzle-orm";
+import createHttpError from "http-errors";
+import HttpCode from "@server/types/HttpCode";
+import { checkOrgAccessPolicy } from "#dynamic/lib/checkOrgAccessPolicy";
+import { getUserOrgRoleIds } from "@server/lib/userOrgRoles";
+import { getFirstString } from "@server/lib/requestParams";
+
+export async function verifyAiBudgetAccess(
+    req: Request,
+    res: Response,
+    next: NextFunction
+) {
+    try {
+        const userId = req.user!.userId;
+        const budgetIdRaw = getFirstString(req.params.budgetId);
+        const budgetId = Number.parseInt(budgetIdRaw ?? "", 10);
+
+        if (!userId) {
+            return next(
+                createHttpError(HttpCode.UNAUTHORIZED, "User not authenticated")
+            );
+        }
+
+        if (Number.isNaN(budgetId)) {
+            return next(
+                createHttpError(HttpCode.BAD_REQUEST, "Invalid budget ID")
+            );
+        }
+
+        const [budget] = await db
+            .select()
+            .from(aiBudgets)
+            .where(eq(aiBudgets.budgetId, budgetId))
+            .limit(1);
+
+        if (!budget) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI budget with ID ${budgetId} not found`
+                )
+            );
+        }
+
+        const orgId = budget.orgId;
+
+        if (!req.userOrg || req.userOrg.orgId !== orgId) {
+            const userOrgRole = await db
+                .select()
+                .from(userOrgs)
+                .where(
+                    and(eq(userOrgs.userId, userId), eq(userOrgs.orgId, orgId))
+                )
+                .limit(1);
+            req.userOrg = userOrgRole[0];
+        }
+
+        if (!req.userOrg) {
+            return next(
+                createHttpError(
+                    HttpCode.FORBIDDEN,
+                    "User does not have access to this organization"
+                )
+            );
+        }
+
+        if (req.orgPolicyAllowed === undefined && req.userOrg.orgId) {
+            const policyCheck = await checkOrgAccessPolicy({
+                orgId: req.userOrg.orgId,
+                userId,
+                session: req.session
+            });
+            req.orgPolicyAllowed = policyCheck.allowed;
+            if (!policyCheck.allowed || policyCheck.error) {
+                return next(
+                    createHttpError(
+                        HttpCode.FORBIDDEN,
+                        "" + (policyCheck.error || "Unknown error")
+                    )
+                );
+            }
+        }
+
+        req.userOrgId = orgId;
+        req.userOrgRoleIds = await getUserOrgRoleIds(req.userOrg.userId, orgId);
+        req.aiBudget = budget;
+
+        return next();
+    } catch (error) {
+        return next(
+            createHttpError(
+                HttpCode.INTERNAL_SERVER_ERROR,
+                "Error verifying AI budget access"
+            )
+        );
+    }
+}

server/middlewares/verifyAiModelAccess.ts

@@ -0,0 +1,107 @@
+import { Request, Response, NextFunction } from "express";
+import { aiModels, aiProviders, db, userOrgs } from "@server/db";
+import { and, eq } from "drizzle-orm";
+import createHttpError from "http-errors";
+import HttpCode from "@server/types/HttpCode";
+import { checkOrgAccessPolicy } from "#dynamic/lib/checkOrgAccessPolicy";
+import { getUserOrgRoleIds } from "@server/lib/userOrgRoles";
+import { getFirstString } from "@server/lib/requestParams";
+
+export async function verifyAiModelAccess(
+    req: Request,
+    res: Response,
+    next: NextFunction
+) {
+    try {
+        const userId = req.user!.userId;
+        const modelIdRaw = getFirstString(req.params.modelId);
+        const modelId = Number.parseInt(modelIdRaw ?? "", 10);
+
+        if (!userId) {
+            return next(
+                createHttpError(HttpCode.UNAUTHORIZED, "User not authenticated")
+            );
+        }
+
+        if (Number.isNaN(modelId)) {
+            return next(
+                createHttpError(HttpCode.BAD_REQUEST, "Invalid model ID")
+            );
+        }
+
+        const [row] = await db
+            .select({
+                model: aiModels,
+                provider: aiProviders
+            })
+            .from(aiModels)
+            .innerJoin(
+                aiProviders,
+                eq(aiModels.providerId, aiProviders.providerId)
+            )
+            .where(eq(aiModels.modelId, modelId))
+            .limit(1);
+
+        if (!row) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI model with ID ${modelId} not found`
+                )
+            );
+        }
+
+        const orgId = row.provider.orgId;
+
+        if (!req.userOrg || req.userOrg.orgId !== orgId) {
+            const userOrgRole = await db
+                .select()
+                .from(userOrgs)
+                .where(
+                    and(eq(userOrgs.userId, userId), eq(userOrgs.orgId, orgId))
+                )
+                .limit(1);
+            req.userOrg = userOrgRole[0];
+        }
+
+        if (!req.userOrg) {
+            return next(
+                createHttpError(
+                    HttpCode.FORBIDDEN,
+                    "User does not have access to this organization"
+                )
+            );
+        }
+
+        if (req.orgPolicyAllowed === undefined && req.userOrg.orgId) {
+            const policyCheck = await checkOrgAccessPolicy({
+                orgId: req.userOrg.orgId,
+                userId,
+                session: req.session
+            });
+            req.orgPolicyAllowed = policyCheck.allowed;
+            if (!policyCheck.allowed || policyCheck.error) {
+                return next(
+                    createHttpError(
+                        HttpCode.FORBIDDEN,
+                        "" + (policyCheck.error || "Unknown error")
+                    )
+                );
+            }
+        }
+
+        req.userOrgId = orgId;
+        req.userOrgRoleIds = await getUserOrgRoleIds(req.userOrg.userId, orgId);
+        req.aiProvider = row.provider;
+        req.aiModel = row.model;
+
+        return next();
+    } catch (error) {
+        return next(
+            createHttpError(
+                HttpCode.INTERNAL_SERVER_ERROR,
+                "Error verifying AI model access"
+            )
+        );
+    }
+}

server/middlewares/verifyAiProviderAccess.ts

@@ -0,0 +1,99 @@
+import { Request, Response, NextFunction } from "express";
+import { aiProviders, db, userOrgs } from "@server/db";
+import { and, eq } from "drizzle-orm";
+import createHttpError from "http-errors";
+import HttpCode from "@server/types/HttpCode";
+import { checkOrgAccessPolicy } from "#dynamic/lib/checkOrgAccessPolicy";
+import { getUserOrgRoleIds } from "@server/lib/userOrgRoles";
+import { getFirstString } from "@server/lib/requestParams";
+
+export async function verifyAiProviderAccess(
+    req: Request,
+    res: Response,
+    next: NextFunction
+) {
+    try {
+        const userId = req.user!.userId;
+        const providerIdRaw = getFirstString(req.params.providerId);
+        const providerId = Number.parseInt(providerIdRaw ?? "", 10);
+
+        if (!userId) {
+            return next(
+                createHttpError(HttpCode.UNAUTHORIZED, "User not authenticated")
+            );
+        }
+
+        if (Number.isNaN(providerId)) {
+            return next(
+                createHttpError(HttpCode.BAD_REQUEST, "Invalid provider ID")
+            );
+        }
+
+        const [provider] = await db
+            .select()
+            .from(aiProviders)
+            .where(eq(aiProviders.providerId, providerId))
+            .limit(1);
+
+        if (!provider) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI provider with ID ${providerId} not found`
+                )
+            );
+        }
+
+        const orgId = provider.orgId;
+
+        if (!req.userOrg || req.userOrg.orgId !== orgId) {
+            const userOrgRole = await db
+                .select()
+                .from(userOrgs)
+                .where(
+                    and(eq(userOrgs.userId, userId), eq(userOrgs.orgId, orgId))
+                )
+                .limit(1);
+            req.userOrg = userOrgRole[0];
+        }
+
+        if (!req.userOrg) {
+            return next(
+                createHttpError(
+                    HttpCode.FORBIDDEN,
+                    "User does not have access to this organization"
+                )
+            );
+        }
+
+        if (req.orgPolicyAllowed === undefined && req.userOrg.orgId) {
+            const policyCheck = await checkOrgAccessPolicy({
+                orgId: req.userOrg.orgId,
+                userId,
+                session: req.session
+            });
+            req.orgPolicyAllowed = policyCheck.allowed;
+            if (!policyCheck.allowed || policyCheck.error) {
+                return next(
+                    createHttpError(
+                        HttpCode.FORBIDDEN,
+                        "" + (policyCheck.error || "Unknown error")
+                    )
+                );
+            }
+        }
+
+        req.userOrgId = orgId;
+        req.userOrgRoleIds = await getUserOrgRoleIds(req.userOrg.userId, orgId);
+        req.aiProvider = provider;
+
+        return next();
+    } catch (error) {
+        return next(
+            createHttpError(
+                HttpCode.INTERNAL_SERVER_ERROR,
+                "Error verifying AI provider access"
+            )
+        );
+    }
+}

server/middlewares/verifyTargetAccess.ts

@@ -1,6 +1,6 @@
 import { Request, Response, NextFunction } from "express";
 import { db } from "@server/db";
-import { resources, targets, userOrgs } from "@server/db";
+import { aiProviders, resources, targets, userOrgs } from "@server/db";
 import { and, eq } from "drizzle-orm";
 import createHttpError from "http-errors";
 import HttpCode from "@server/types/HttpCode";
@@ -25,9 +25,7 @@ export async function verifyTargetAccess(
     }
 
     if (isNaN(targetId)) {
-        return next(
-            createHttpError(HttpCode.BAD_REQUEST, "Invalid organization ID")
-        );
+        return next(createHttpError(HttpCode.BAD_REQUEST, "Invalid target ID"));
     }
 
     const target = await db
@@ -45,73 +43,88 @@ export async function verifyTargetAccess(
         );
     }
 
-    const resourceId = target[0].resourceId;
+    const { resourceId, providerId } = target[0];
 
-    if (!resourceId) {
+    if ((!resourceId && !providerId) || (resourceId && providerId)) {
         return next(
             createHttpError(
                 HttpCode.INTERNAL_SERVER_ERROR,
-                `Target with ID ${targetId} does not have a resource ID`
+                `Target with ID ${targetId} has invalid ownership`
             )
         );
     }
 
     try {
-        const resource = await db
-            .select()
-            .from(resources)
-            .where(eq(resources.resourceId, resourceId!))
-            .limit(1);
+        let orgId: string;
 
-        if (resource.length === 0) {
-            return next(
-                createHttpError(
-                    HttpCode.NOT_FOUND,
-                    `Resource with ID ${resourceId} not found`
-                )
-            );
-        }
+        if (resourceId) {
+            const [resource] = await db
+                .select()
+                .from(resources)
+                .where(eq(resources.resourceId, resourceId))
+                .limit(1);
 
-        if (!resource[0].orgId) {
-            return next(
-                createHttpError(
-                    HttpCode.INTERNAL_SERVER_ERROR,
-                    `resource with ID ${resourceId} does not have an organization ID`
-                )
-            );
+            if (!resource) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `Resource with ID ${resourceId} not found`
+                    )
+                );
+            }
+
+            if (!resource.orgId) {
+                return next(
+                    createHttpError(
+                        HttpCode.INTERNAL_SERVER_ERROR,
+                        `Resource with ID ${resourceId} does not have an organization ID`
+                    )
+                );
+            }
+
+            orgId = resource.orgId;
+        } else {
+            const [provider] = await db
+                .select()
+                .from(aiProviders)
+                .where(eq(aiProviders.providerId, providerId!))
+                .limit(1);
+
+            if (!provider) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `AI provider with ID ${providerId} not found`
+                    )
+                );
+            }
+
+            orgId = provider.orgId;
         }
 
         if (!req.userOrg) {
-            const res = await db
+            const userOrgResult = await db
                 .select()
                 .from(userOrgs)
                 .where(
-                    and(
-                        eq(userOrgs.userId, userId),
-                        eq(userOrgs.orgId, resource[0].orgId)
-                    )
+                    and(eq(userOrgs.userId, userId), eq(userOrgs.orgId, orgId))
                 );
-            req.userOrg = res[0];
+            req.userOrg = userOrgResult[0];
         }
 
-        if (!req.userOrg) {
-            next(
+        if (!req.userOrg || req.userOrg.orgId !== orgId) {
+            return next(
                 createHttpError(
                     HttpCode.FORBIDDEN,
                     "User does not have access to this organization"
                 )
             );
-        } else {
-            req.userOrgRoleIds = await getUserOrgRoleIds(
-                req.userOrg.userId,
-                resource[0].orgId!
-            );
-            req.userOrgId = resource[0].orgId!;
         }
 
-        const orgId = req.userOrg.orgId;
+        req.userOrgRoleIds = await getUserOrgRoleIds(req.userOrg.userId, orgId);
+        req.userOrgId = orgId;
 
-        if (req.orgPolicyAllowed === undefined && orgId) {
+        if (req.orgPolicyAllowed === undefined) {
             const policyCheck = await checkOrgAccessPolicy({
                 orgId,
                 userId,
@@ -128,22 +141,24 @@ export async function verifyTargetAccess(
             }
         }
 
-        const resourceAllowed = await canUserAccessResource({
-            userId,
-            resourceId,
-            roleIds: req.userOrgRoleIds ?? []
-        });
+        if (resourceId) {
+            const resourceAllowed = await canUserAccessResource({
+                userId,
+                resourceId,
+                roleIds: req.userOrgRoleIds ?? []
+            });
 
-        if (!resourceAllowed) {
-            return next(
-                createHttpError(
-                    HttpCode.FORBIDDEN,
-                    "User does not have access to this resource"
-                )
-            );
+            if (!resourceAllowed) {
+                return next(
+                    createHttpError(
+                        HttpCode.FORBIDDEN,
+                        "User does not have access to this resource"
+                    )
+                );
+            }
         }
 
-        next();
+        return next();
     } catch (e) {
         return next(
             createHttpError(

server/openApi.ts

@@ -28,7 +28,10 @@ export enum OpenAPITags {
     HealthCheck = "Health Check",
     PublicResourcePolicyLegacy = "Public Resource Policy (Legacy)",
     PublicResourceLegacy = "Public Resource (Legacy)",
-    PrivateResourceLegacy = "Private Resource (Legacy)"
+    PrivateResourceLegacy = "Private Resource (Legacy)",
+    AiProvider = "AI Provider",
+    AiModel = "AI Model",
+    AiBudget = "AI Budget"
 }
 
 // Order here controls the order tags are displayed in Swagger UI

server/private/lib/exitNodes/exitNodes.ts

@@ -25,7 +25,6 @@ import {
     Transaction
 } from "@server/db";
 import logger from "@server/logger";
-import { ExitNodePingResult } from "@server/routers/newt";
 import { eq, and, or, ne, isNull, inArray } from "drizzle-orm";
 import axios from "axios";
 import config from "../config";
@@ -330,6 +329,16 @@ export async function listExitNodes(
     return exitNodesList;
 }
 
+export type ExitNodePingResult = {
+    exitNodeId: number;
+    latencyMs: number;
+    weight: number;
+    error?: string;
+    exitNodeName: string;
+    endpoint: string;
+    wasPreviouslyConnected: boolean;
+};
+
 /**
  * Selects the most suitable exit node from a list of ping results.
  *

server/private/lib/traefik/getTraefikConfig.ts

@@ -18,6 +18,7 @@ import {
     domains,
     exitNodes,
     loginPage,
+    SiteResource,
     targetHealthCheck
 } from "@server/db";
 import {
@@ -40,7 +41,10 @@ import {
     siteNetworks,
     siteResources,
     Target,
-    targets
+    targets,
+    aiProviders,
+    resourceAiProviders,
+    siteResourceAiProviders
 } from "@server/db";
 import {
     sanitize,
@@ -54,6 +58,7 @@ import {
     getValidCertificatesForDomains
 } from "#private/lib/certificates";
 import { build } from "@server/build";
+import regionalCache from "#private/lib/cache";
 
 const redirectHttpsMiddlewareName = "redirect-to-https";
 const redirectToRootMiddlewareName = "redirect-to-root";
@@ -87,8 +92,24 @@ export async function getTraefikConfig(
     generateLoginPageRouters = false,
     allowRawResources = true,
     maintenancePageUiUrl: string | null = null,
-    browserGatewayUiUrl: string | null = null
+    browserGatewayUiUrl: string | null = null,
+    aiGatewayUrl: string | null = null
 ): Promise<any> {
+    // Get the exit node but cache it for 5 minutes to avoid hitting the DB too often
+    const exitNodeCacheKey = `exitNode:${exitNodeId}`;
+    let exitNode =
+        await regionalCache.get<typeof exitNodes.$inferSelect>(
+            exitNodeCacheKey
+        );
+    if (!exitNode) {
+        [exitNode] = await db
+            .select()
+            .from(exitNodes)
+            .where(eq(exitNodes.exitNodeId, exitNodeId))
+            .limit(1);
+        await regionalCache.set(exitNodeCacheKey, exitNode, 300);
+    }
+
     // Get resources with their targets and sites in a single optimized query
     // Start from sites on this exit node, then join to targets and resources
     const resourcesWithTargetsAndSites = await db
@@ -136,7 +157,7 @@ export async function getTraefikConfig(
             siteId: sites.siteId,
             siteType: sites.type,
             siteOnline: sites.online,
-            subnet: sites.subnet,
+            subnet: sites.exitNodeSubnet,
             exitNodeId: sites.exitNodeId,
             // Namespace
             domainNamespaceId: domainNamespaces.domainNamespaceId,
@@ -361,7 +382,7 @@ export async function getTraefikConfig(
     let siteResourcesWithFullDomain: {
         siteResourceId: number;
         fullDomain: string | null;
-        mode: "http" | "host" | "cidr" | "ssh";
+        mode: SiteResource["mode"];
     }[] = [];
     if (
         build == "enterprise" &&
@@ -386,7 +407,7 @@ export async function getTraefikConfig(
                 and(
                     eq(siteResources.enabled, true),
                     isNotNull(siteResources.fullDomain),
-                    eq(siteResources.mode, "http"),
+                    eq(siteResources.mode, "http"), // important so we dont double get the inference siteResources below
                     eq(siteResources.ssl, true),
                     eq(sites.exitNodeId, exitNodeId),
                     inArray(sites.type, siteTypes)
@@ -394,6 +415,56 @@ export async function getTraefikConfig(
             );
     }
 
+    // Inference-mode resources/siteResources have no targets/sites/network
+    // (their "backend" is the central AI gateway, not something on a site),
+    // so they can't be reached via the joins above - query them separately
+    // and include them on every exit node.
+    const inferenceResources = await db
+        .selectDistinct({
+            resourceId: resources.resourceId,
+            fullDomain: resources.fullDomain,
+            ssl: resources.ssl,
+            subdomain: resources.subdomain,
+            domainId: resources.domainId,
+            enabled: resources.enabled,
+            wildcard: resources.wildcard,
+            domainCertResolver: domains.certResolver,
+            preferWildcardCert: domains.preferWildcardCert
+        })
+        .from(resources)
+        .innerJoin(
+            resourceAiProviders,
+            eq(resources.resourceId, resourceAiProviders.resourceId)
+        )
+        .innerJoin(
+            aiProviders,
+            eq(resourceAiProviders.providerId, aiProviders.providerId)
+        )
+        .leftJoin(domains, eq(domains.domainId, resources.domainId))
+        .where(
+            and(
+                eq(resources.mode, "inference"),
+                eq(resources.enabled, true),
+                eq(aiProviders.enabled, true)
+            )
+        );
+
+    const siteResourcesInference = await db
+        .selectDistinct({
+            siteResourceId: siteResources.siteResourceId,
+            fullDomain: siteResources.fullDomain,
+            ssl: siteResources.ssl,
+            enabled: siteResources.enabled
+        })
+        .from(siteResources)
+        .where(
+            and(
+                eq(siteResources.mode, "inference"),
+                eq(siteResources.enabled, true),
+                isNotNull(siteResources.fullDomain)
+            )
+        );
+
     let validCerts: CertificateResult[] = [];
     if (privateConfig.getRawPrivateConfig().flags.use_pangolin_dns) {
         // create a list of all domains to get certs for
@@ -415,6 +486,17 @@ export async function getTraefikConfig(
                 domains.add(bgResource.fullDomain);
             }
         }
+        // Include inference resource/siteResource domains
+        for (const ir of inferenceResources) {
+            if (ir.enabled && ir.ssl && ir.fullDomain) {
+                domains.add(ir.fullDomain);
+            }
+        }
+        for (const sr of siteResourcesInference) {
+            if (sr.enabled && sr.ssl && sr.fullDomain) {
+                domains.add(sr.fullDomain);
+            }
+        }
         // get the valid certs for these domains
         validCerts = await getValidCertificatesForDomains(domains, true); // we are caching here because this is called often
         // logger.debug(`Valid certs for domains: ${JSON.stringify(validCerts)}`);
@@ -1461,6 +1543,244 @@ export async function getTraefikConfig(
         }
     }
 
+    if (aiGatewayUrl) {
+        // The AI gateway may live on a different host than the inference
+        // resource itself (e.g. a remote exit node forwarding to the
+        // central dashboard over a tunnel). passHostHeader would forward
+        // the resource's own Host, which that external host won't
+        // recognize, so we pin the Host header to the gateway's own host
+        // and smuggle the original resource host through in "p-host"
+        // instead (same pattern as the maintenance-page routes above).
+        let aiGatewayHost: string | undefined;
+        try {
+            aiGatewayHost = new URL(aiGatewayUrl).host;
+        } catch {
+            aiGatewayHost = undefined;
+        }
+
+        // The p-host smuggling above is only necessary when the AI gateway
+        // is overridden to a different host than the resource's own. In the
+        // default case, leave the Host header untouched so it's visible on
+        // the other end.
+        const aiGatewayOverride =
+            config.getRawConfig().server.ai_gateway_override;
+
+        // Public inference resources: same TLS/cert-resolver handling as
+        // plain http-mode resources, but the service points at the AI
+        // gateway instead of any real backend targets.
+        for (const ir of inferenceResources) {
+            if (!ir.enabled) continue;
+            if (!ir.domainId || !ir.fullDomain) continue;
+
+            if (!config_output.http.routers) config_output.http.routers = {};
+            if (!config_output.http.services) config_output.http.services = {};
+
+            const fullDomain = ir.fullDomain;
+            const irKey = `inference-r${ir.resourceId}`;
+            const routerName = `${irKey}-router`;
+            const serviceName = `${irKey}-service`;
+
+            let rule: string;
+            if (ir.wildcard && fullDomain.startsWith("*.")) {
+                const escaped = fullDomain.slice(2).replace(/\./g, "\\.");
+                rule = `HostRegexp(\`^[^.]+\\.${escaped}$\`)`;
+            } else {
+                rule = `Host(\`${fullDomain}\`)`;
+            }
+
+            let tls: any = {};
+            if (!privateConfig.getRawPrivateConfig().flags.use_pangolin_dns) {
+                const domainParts = fullDomain.split(".");
+                let wildCard;
+                if (domainParts.length <= 2) {
+                    wildCard = `*.${domainParts.join(".")}`;
+                } else {
+                    wildCard = `*.${domainParts.slice(1).join(".")}`;
+                }
+                if (!ir.subdomain) {
+                    wildCard = fullDomain;
+                }
+
+                const globalDefaultResolver =
+                    config.getRawConfig().traefik.cert_resolver;
+                const globalDefaultPreferWildcard =
+                    config.getRawConfig().traefik.prefer_wildcard_cert;
+                const resolverName = ir.domainCertResolver
+                    ? ir.domainCertResolver.trim()
+                    : globalDefaultResolver;
+                const preferWildcard =
+                    ir.preferWildcardCert !== undefined &&
+                    ir.preferWildcardCert !== null
+                        ? ir.preferWildcardCert
+                        : globalDefaultPreferWildcard;
+
+                tls = {
+                    certResolver: resolverName,
+                    ...(preferWildcard ? { domains: [{ main: wildCard }] } : {})
+                };
+            } else {
+                const matchingCert = validCerts.find(
+                    (cert) => cert.queriedDomain === fullDomain
+                );
+                if (!matchingCert) {
+                    logger.debug(
+                        `No matching certificate found for inference resource domain: ${fullDomain}`
+                    );
+                    continue;
+                }
+            }
+
+            const additionalMiddlewares =
+                config.getRawConfig().traefik.additional_middlewares || [];
+            const routerMiddlewares = [badgerMiddlewareName];
+
+            if (aiGatewayOverride) {
+                const irHeadersMiddlewareName = `${irKey}-headers-middleware`;
+                config_output.http.middlewares[irHeadersMiddlewareName] = {
+                    headers: {
+                        customRequestHeaders: {
+                            ...(aiGatewayHost ? { Host: aiGatewayHost } : {}),
+                            "p-host": fullDomain
+                        }
+                    }
+                };
+                routerMiddlewares.push(irHeadersMiddlewareName);
+            }
+
+            routerMiddlewares.push(...additionalMiddlewares);
+
+            if (ir.ssl) {
+                config_output.http.routers[routerName + "-redirect"] = {
+                    entryPoints: [
+                        config.getRawConfig().traefik.http_entrypoint
+                    ],
+                    middlewares: [redirectHttpsMiddlewareName],
+                    service: serviceName,
+                    rule,
+                    priority: 100
+                };
+            }
+
+            config_output.http.routers[routerName] = {
+                entryPoints: [
+                    ir.ssl
+                        ? config.getRawConfig().traefik.https_entrypoint
+                        : config.getRawConfig().traefik.http_entrypoint
+                ],
+                middlewares: routerMiddlewares,
+                service: serviceName,
+                rule,
+                priority: 100,
+                ...(ir.ssl ? { tls } : {})
+            };
+
+            config_output.http.services[serviceName] = {
+                loadBalancer: {
+                    servers: [{ url: aiGatewayUrl }]
+                }
+            };
+        }
+
+        // Private (siteResource) inference resources: routed by their alias
+        // instead of a public fullDomain, and deliberately WITHOUT the
+        // badger middleware - no per-user auth/policy stack exists for
+        // siteResources today (see plan doc), so gating here is
+        // reachability-only for now.
+        for (const sr of siteResourcesInference) {
+            if (!sr.enabled || !sr.fullDomain) continue;
+
+            if (!config_output.http.routers) config_output.http.routers = {};
+            if (!config_output.http.services) config_output.http.services = {};
+
+            const fullDomain = sr.fullDomain;
+            const srKey = `inference-sr${sr.siteResourceId}`;
+            const routerName = `${srKey}-router`;
+            const serviceName = `${srKey}-service`;
+            const rule = `Host(\`${fullDomain}\`) && ClientIP(\`${exitNode.address}\`)`; // restrict to coming from the exit node ip range that the client is connected to
+
+            let tls: any = {};
+            if (!privateConfig.getRawPrivateConfig().flags.use_pangolin_dns) {
+                const domainParts = fullDomain.split(".");
+                const wildCard =
+                    domainParts.length <= 2
+                        ? `*.${domainParts.join(".")}`
+                        : `*.${domainParts.slice(1).join(".")}`;
+
+                const globalDefaultResolver =
+                    config.getRawConfig().traefik.cert_resolver;
+                const globalDefaultPreferWildcard =
+                    config.getRawConfig().traefik.prefer_wildcard_cert;
+
+                tls = {
+                    certResolver: globalDefaultResolver,
+                    ...(globalDefaultPreferWildcard
+                        ? { domains: [{ main: wildCard }] }
+                        : {})
+                };
+            } else {
+                const matchingCert = validCerts.find(
+                    (cert) => cert.queriedDomain === fullDomain
+                );
+                if (!matchingCert) {
+                    logger.debug(
+                        `No matching certificate found for inference siteResource fullDomain: ${fullDomain}`
+                    );
+                    continue;
+                }
+            }
+
+            const additionalMiddlewares =
+                config.getRawConfig().traefik.additional_middlewares || [];
+            const routerMiddlewares: string[] = [];
+
+            if (aiGatewayOverride) {
+                const srHeadersMiddlewareName = `${srKey}-headers-middleware`;
+                config_output.http.middlewares[srHeadersMiddlewareName] = {
+                    headers: {
+                        customRequestHeaders: {
+                            ...(aiGatewayHost ? { Host: aiGatewayHost } : {}),
+                            "p-host": fullDomain
+                        }
+                    }
+                };
+                routerMiddlewares.push(srHeadersMiddlewareName);
+            }
+
+            routerMiddlewares.push(...additionalMiddlewares);
+
+            if (sr.ssl) {
+                config_output.http.routers[routerName + "-redirect"] = {
+                    entryPoints: [
+                        config.getRawConfig().traefik.http_entrypoint
+                    ],
+                    middlewares: [redirectHttpsMiddlewareName],
+                    service: serviceName,
+                    rule,
+                    priority: 100
+                };
+            }
+
+            config_output.http.routers[routerName] = {
+                entryPoints: [
+                    sr.ssl
+                        ? config.getRawConfig().traefik.https_entrypoint
+                        : config.getRawConfig().traefik.http_entrypoint
+                ],
+                middlewares: routerMiddlewares,
+                service: serviceName,
+                rule,
+                priority: 100,
+                ...(sr.ssl ? { tls } : {})
+            };
+
+            config_output.http.services[serviceName] = {
+                loadBalancer: {
+                    servers: [{ url: aiGatewayUrl }]
+                }
+            };
+        }
+    }
+
     if (generateLoginPageRouters) {
         const exitNodeLoginPages = await db
             .select({

server/private/routers/hybrid.ts

@@ -351,6 +351,7 @@ hybridRouter.get(
         }
 
         const pangolinUIUrl = config.getRawConfig().app.dashboard_url; // points to the dashboard to serve from there
+        const aiGatewayUrl = `${config.getRawConfig().app.dashboard_url}/api/v1/ai-gateway`;
 
         try {
             const traefikConfig = await getTraefikConfig(
@@ -360,7 +361,8 @@ hybridRouter.get(
                 false, // Dont include login pages,
                 true, // allow raw resources
                 pangolinUIUrl, // dont generate maintenance page
-                pangolinUIUrl // generate browser gateway targets
+                pangolinUIUrl, // generate browser gateway targets
+                aiGatewayUrl
             );
 
             return response(res, {

server/private/routers/re-key/reGenerateSiteSecret.ts

@@ -178,7 +178,7 @@ export async function reGenerateSiteSecret(
                     );
                 }
 
-                if (site.exitNodeId && site.subnet) {
+                if (site.exitNodeId && site.exitNodeSubnet) {
                     await deletePeer(site.exitNodeId, site.pubKey!); // the old pubkey
                     await addPeer(site.exitNodeId, {
                         publicKey: pubKey,

server/routers/aiBudget/README.md

@@ -0,0 +1,126 @@
+# AI Budget API
+
+Public/OSS CRUD entity (`server/routers/aiBudget/`, not enterprise-gated).
+Table: `aiBudgets` in `server/db/{pg,sqlite}/schema/schema.ts`, type `AiBudget`.
+
+## What a budget is
+
+A row is a spend/usage cap of `amount` `unit` (`usd` | `tokens`) per `period`
+(`hourly` | `daily` | `weekly` | `monthly` | `yearly` | `lifetime`), with
+`enforcement` (`hard` | `soft`) and an `enabled` flag.
+
+Every budget belongs to an org (`orgId`, required) and is optionally further
+scoped to **exactly one** of:
+
+- `providerId` → an `aiProviders` row
+- `modelId` → an `aiModels` row
+- `resourceId` → a `resources` row
+- `siteResourceId` → a `siteResources` row
+- `roleId` → a `roles` row
+
+If none of those five are set, the budget is **org-wide**. Setting more than
+one at once is rejected by `validation.ts`'s `refineBudgetScopeFields`
+(`400`, "Only one of providerId, modelId, resourceId, siteResourceId, or
+roleId may be set on a budget").
+
+## Uniqueness / conflict rule
+
+A given scope (one specific provider, or model, or resource, or site
+resource, or role, or "org-wide") may have **multiple** budgets, but at most
+**one per `(unit, period)` combination** — e.g. one `weekly`/`usd` budget and
+one `hourly`/`usd` budget can coexist on the same provider, but two
+`weekly`/`usd` budgets cannot. This is enforced at two levels:
+
+- DB: composite `unique` constraints in both schema files —
+  `ai_budget_provider_uniq (providerId, unit, period)`,
+  `ai_budget_model_uniq (modelId, unit, period)`,
+  `ai_budget_resource_uniq (resourceId, unit, period)`,
+  `ai_budget_site_resource_uniq (siteResourceId, unit, period)`,
+  `ai_budget_role_uniq (roleId, unit, period)`. (NULL scope columns never
+  collide under a plain unique index, so this does *not* cover the org-wide
+  case — see next bullet.)
+- App: `createAiBudget`/`updateAiBudget` both run an explicit pre-check
+  query keyed on `(scopeCondition, unit, period)` before insert/update,
+  where `scopeCondition` is `eq(<scopeColumn>, id)` for whichever scope
+  field is set, or — when none is set — `orgId = X AND` all five scope
+  columns `IS NULL`, so org-wide budgets get the same one-per-`(unit,
+  period)` guarantee even though the DB constraint can't express it.
+  Violating this returns `409` with
+  `` `A ${period} ${unit} budget already exists for this scope` ``.
+
+Because only one row can ever exist for a given `(scope, unit, period)`,
+there is no separate check needed to prevent a `hard` and a `soft` budget
+from coexisting on the same `(scope, unit, period)` — the conflict check
+above already blocks the second row regardless of its `enforcement` value.
+
+On `updateAiBudget`, the conflict/ownership checks are run against the
+**merged** next-state (existing row's scope/unit/period overlaid with
+whatever the request body changes), not just the fields present in the
+body — so e.g. changing only `unit` on a budget that already has
+`providerId` set re-validates against that provider's other budgets at the
+new unit.
+
+## Ownership validation
+
+`providerId`/`modelId`/`resourceId`/`siteResourceId`/`roleId` are validated
+to belong to the same `orgId` as the budget (`modelId` via an
+`aiModels ⋈ aiProviders` join, since `aiModels` has no `orgId` column
+directly). A mismatch returns `404`, not `403` — this matches how the
+sibling `aiProvider`/`aiModel` routers report cross-org references.
+
+## Routes
+
+All under `server/routers/external.ts`, registered right after the
+`aiProvider`/`aiModel` block. `PUT` = create, `POST` = update (repo
+convention, not standard REST).
+
+| Method | Path | Middleware | Action | Handler |
+|---|---|---|---|---|
+| PUT | `/org/:orgId/ai-budget` | `verifyOrgAccess` | `createAiBudget` | `createAiBudget` |
+| GET | `/org/:orgId/ai-budgets` | `verifyOrgAccess` | `listAiBudgets` | `listAiBudgets` (paginated) |
+| GET | `/ai-budget/:budgetId` | `verifyAiBudgetAccess` | `getAiBudget` | `getAiBudget` |
+| POST | `/ai-budget/:budgetId` | `verifyAiBudgetAccess` | `updateAiBudget` | `updateAiBudget` |
+| DELETE | `/ai-budget/:budgetId` | `verifyAiBudgetAccess` | `deleteAiBudget` | `deleteAiBudget` |
+| GET | `/ai-provider/:providerId/ai-budgets` | `verifyAiProviderAccess` | `listAiBudgets` | `listAiBudgetsForProvider` |
+| GET | `/ai-model/:modelId/ai-budgets` | `verifyAiModelAccess` | `listAiBudgets` | `listAiBudgetsForModel` |
+| GET | `/resource/:resourceId/ai-budgets` | `verifyResourceAccess` | `listAiBudgets` | `listAiBudgetsForResource` |
+| GET | `/site-resource/:siteResourceId/ai-budgets` | `verifySiteResourceAccess` | `listAiBudgets` | `listAiBudgetsForSiteResource` |
+| GET | `/role/:roleId/ai-budgets` | `verifyRoleAccess` | `listAiBudgets` | `listAiBudgetsForRole` |
+
+The five scope-filtered `GET .../ai-budgets` routes intentionally reuse the
+single `ActionsEnum.listAiBudgets` action rather than getting one action
+each — access control is already fully handled by the entity-specific
+middleware (a user who can see the provider/resource/etc. can see its
+budgets), so per-scope actions would just be enum bloat. They also skip
+pagination (unlike the org-wide list) since a single entity realistically
+has only a handful of `(unit, period)` budgets — response shape is a flat
+`{ budgets: AiBudget[] }` (`ListAiBudgetsByScopeResponse`), not
+`PaginatedResponse`.
+
+`verifyAiBudgetAccess` (`server/middlewares/verifyAiBudgetAccess.ts`) loads
+the budget by `budgetId`, resolves its `orgId` directly off the row (no
+join needed, unlike `verifyAiModelAccess`), and stashes it on
+`req.aiBudget` so `getAiBudget`/`updateAiBudget` can skip a re-fetch.
+
+## Request/response shapes
+
+- Create body: `providerId?`, `modelId?`, `resourceId?`, `siteResourceId?`,
+  `roleId?` (all `number`, mutually exclusive), `amount` (positive
+  `number`, required), `unit` (required), `period` (default `"monthly"`),
+  `enforcement` (default `"hard"`), `enabled?` (default `true`).
+- Update body: same fields, all optional; the five scope fields are
+  `nullable().optional()` so a client can explicitly send `null` to clear
+  a scope (turning a scoped budget into an org-wide one).
+- All five CRUD responses wrap a single `budget: AiBudget` (or
+  `budgets: AiBudget[]` + `pagination` for the org-wide list). No public/
+  private mapper exists for `AiBudget` — unlike `AiProvider`, there's no
+  secret field to strip, so the raw DB row is returned as-is.
+
+## Not yet migrated
+
+Schema changes here (composite unique constraints) were made directly in
+`schema.ts` without hand-writing a `server/migrations/*.sql` file — this
+repo's CI (`.github/workflows/test.yml`) runs `drizzle-kit generate`
+against `schema.ts` fresh, and other recent schema-only commits (e.g. "Remove
+budget periods") follow the same pattern of not committing a matching
+migration by hand.

server/routers/aiBudget/createAiBudget.ts

@@ -0,0 +1,265 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import {
+    aiBudgets,
+    aiModels,
+    aiProviders,
+    db,
+    resources,
+    roles,
+    siteResources
+} from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { and, eq, isNull } from "drizzle-orm";
+import type { CreateOrEditAiBudgetResponse } from "@server/routers/aiBudget/types";
+import {
+    aiBudgetEnforcementSchema,
+    aiBudgetPeriodSchema,
+    aiBudgetUnitSchema,
+    refineBudgetScopeFields
+} from "@server/routers/aiBudget/validation";
+
+const paramsSchema = z.strictObject({
+    orgId: z.string().nonempty()
+});
+
+const bodySchema = z
+    .strictObject({
+        providerId: z.coerce.number().int().positive().optional(),
+        modelId: z.coerce.number().int().positive().optional(),
+        resourceId: z.coerce.number().int().positive().optional(),
+        siteResourceId: z.coerce.number().int().positive().optional(),
+        roleId: z.coerce.number().int().positive().optional(),
+        amount: z.number().positive(),
+        unit: aiBudgetUnitSchema,
+        period: aiBudgetPeriodSchema.optional().default("monthly"),
+        enforcement: aiBudgetEnforcementSchema.optional().default("hard"),
+        enabled: z.boolean().optional()
+    })
+    .superRefine((data, ctx) => refineBudgetScopeFields(data, ctx));
+
+registry.registerPath({
+    method: "put",
+    path: "/org/{orgId}/ai-budget",
+    description: "Create an AI budget for an organization.",
+    tags: [OpenAPITags.AiBudget],
+    request: {
+        params: paramsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: bodySchema
+                }
+            }
+        }
+    },
+    responses: {
+        201: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function createAiBudget(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const parsedBody = bodySchema.safeParse(req.body);
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { orgId } = parsedParams.data;
+        const {
+            providerId,
+            modelId,
+            resourceId,
+            siteResourceId,
+            roleId,
+            amount,
+            unit,
+            period,
+            enforcement,
+            enabled
+        } = parsedBody.data;
+
+        if (providerId !== undefined) {
+            const [provider] = await db
+                .select({ orgId: aiProviders.orgId })
+                .from(aiProviders)
+                .where(eq(aiProviders.providerId, providerId))
+                .limit(1);
+            if (!provider || provider.orgId !== orgId) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `AI provider with ID ${providerId} not found in this organization`
+                    )
+                );
+            }
+        }
+
+        if (modelId !== undefined) {
+            const [model] = await db
+                .select({ orgId: aiProviders.orgId })
+                .from(aiModels)
+                .innerJoin(
+                    aiProviders,
+                    eq(aiModels.providerId, aiProviders.providerId)
+                )
+                .where(eq(aiModels.modelId, modelId))
+                .limit(1);
+            if (!model || model.orgId !== orgId) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `AI model with ID ${modelId} not found in this organization`
+                    )
+                );
+            }
+        }
+
+        if (resourceId !== undefined) {
+            const [resource] = await db
+                .select({ orgId: resources.orgId })
+                .from(resources)
+                .where(eq(resources.resourceId, resourceId))
+                .limit(1);
+            if (!resource || resource.orgId !== orgId) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `Resource with ID ${resourceId} not found in this organization`
+                    )
+                );
+            }
+        }
+
+        if (siteResourceId !== undefined) {
+            const [siteResource] = await db
+                .select({ orgId: siteResources.orgId })
+                .from(siteResources)
+                .where(eq(siteResources.siteResourceId, siteResourceId))
+                .limit(1);
+            if (!siteResource || siteResource.orgId !== orgId) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `Site resource with ID ${siteResourceId} not found in this organization`
+                    )
+                );
+            }
+        }
+
+        if (roleId !== undefined) {
+            const [role] = await db
+                .select({ orgId: roles.orgId })
+                .from(roles)
+                .where(eq(roles.roleId, roleId))
+                .limit(1);
+            if (!role || role.orgId !== orgId) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `Role with ID ${roleId} not found in this organization`
+                    )
+                );
+            }
+        }
+
+        const scopeCondition =
+            providerId !== undefined
+                ? eq(aiBudgets.providerId, providerId)
+                : modelId !== undefined
+                  ? eq(aiBudgets.modelId, modelId)
+                  : resourceId !== undefined
+                    ? eq(aiBudgets.resourceId, resourceId)
+                    : siteResourceId !== undefined
+                      ? eq(aiBudgets.siteResourceId, siteResourceId)
+                      : roleId !== undefined
+                        ? eq(aiBudgets.roleId, roleId)
+                        : and(
+                              eq(aiBudgets.orgId, orgId),
+                              isNull(aiBudgets.providerId),
+                              isNull(aiBudgets.modelId),
+                              isNull(aiBudgets.resourceId),
+                              isNull(aiBudgets.siteResourceId),
+                              isNull(aiBudgets.roleId)
+                          );
+
+        const [existing] = await db
+            .select({ budgetId: aiBudgets.budgetId })
+            .from(aiBudgets)
+            .where(
+                and(
+                    scopeCondition,
+                    eq(aiBudgets.unit, unit),
+                    eq(aiBudgets.period, period)
+                )
+            )
+            .limit(1);
+        if (existing) {
+            return next(
+                createHttpError(
+                    HttpCode.CONFLICT,
+                    `A ${period} ${unit} budget already exists for this scope`
+                )
+            );
+        }
+
+        const now = Date.now();
+        const [budget] = await db
+            .insert(aiBudgets)
+            .values({
+                orgId,
+                providerId: providerId ?? null,
+                modelId: modelId ?? null,
+                resourceId: resourceId ?? null,
+                siteResourceId: siteResourceId ?? null,
+                roleId: roleId ?? null,
+                amount,
+                unit,
+                period,
+                enforcement,
+                enabled: enabled ?? true,
+                createdAt: now,
+                updatedAt: now
+            })
+            .returning();
+
+        return response<CreateOrEditAiBudgetResponse>(res, {
+            data: { budget },
+            success: true,
+            error: false,
+            message: "AI budget created successfully",
+            status: HttpCode.CREATED
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiBudget/deleteAiBudget.ts

@@ -0,0 +1,79 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiBudgets, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { eq } from "drizzle-orm";
+
+const paramsSchema = z.strictObject({
+    budgetId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "delete",
+    path: "/ai-budget/{budgetId}",
+    description: "Delete an AI budget.",
+    tags: [OpenAPITags.AiBudget],
+    request: {
+        params: paramsSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function deleteAiBudget(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { budgetId } = parsedParams.data;
+
+        const [existing] = await db
+            .select({ budgetId: aiBudgets.budgetId })
+            .from(aiBudgets)
+            .where(eq(aiBudgets.budgetId, budgetId))
+            .limit(1);
+
+        if (!existing) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI budget with ID ${budgetId} not found`
+                )
+            );
+        }
+
+        await db.delete(aiBudgets).where(eq(aiBudgets.budgetId, budgetId));
+
+        return response(res, {
+            data: null,
+            success: true,
+            error: false,
+            message: "AI budget deleted successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiBudget/getAiBudget.ts

@@ -0,0 +1,81 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiBudgets, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { eq } from "drizzle-orm";
+import type { GetAiBudgetResponse } from "@server/routers/aiBudget/types";
+
+const paramsSchema = z.strictObject({
+    budgetId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "get",
+    path: "/ai-budget/{budgetId}",
+    description: "Get an AI budget by ID.",
+    tags: [OpenAPITags.AiBudget],
+    request: {
+        params: paramsSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function getAiBudget(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { budgetId } = parsedParams.data;
+
+        const [budget] =
+            req.aiBudget && req.aiBudget.budgetId === budgetId
+                ? [req.aiBudget]
+                : await db
+                      .select()
+                      .from(aiBudgets)
+                      .where(eq(aiBudgets.budgetId, budgetId))
+                      .limit(1);
+
+        if (!budget) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI budget with ID ${budgetId} not found`
+                )
+            );
+        }
+
+        return response<GetAiBudgetResponse>(res, {
+            data: { budget },
+            success: true,
+            error: false,
+            message: "AI budget retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiBudget/index.ts

@@ -0,0 +1,11 @@
+export * from "./createAiBudget";
+export * from "./listAiBudgets";
+export * from "./listAiBudgetsForProvider";
+export * from "./listAiBudgetsForModel";
+export * from "./listAiBudgetsForResource";
+export * from "./listAiBudgetsForSiteResource";
+export * from "./listAiBudgetsForRole";
+export * from "./getAiBudget";
+export * from "./updateAiBudget";
+export * from "./deleteAiBudget";
+export * from "./types";

server/routers/aiBudget/listAiBudgets.ts

@@ -0,0 +1,130 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiBudgets, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { asc, eq } from "drizzle-orm";
+import type { ListAiBudgetsResponse } from "@server/routers/aiBudget/types";
+
+const paramsSchema = z.strictObject({
+    orgId: z.string().nonempty()
+});
+
+const listSchema = z.object({
+    pageSize: z.coerce
+        .number<string>()
+        .int()
+        .positive()
+        .optional()
+        .catch(20)
+        .default(20)
+        .openapi({
+            type: "integer",
+            default: 20,
+            description: "Number of items per page"
+        }),
+    page: z.coerce
+        .number<string>()
+        .int()
+        .min(0)
+        .optional()
+        .catch(1)
+        .default(1)
+        .openapi({
+            type: "integer",
+            default: 1,
+            description: "Page number to retrieve"
+        })
+});
+
+registry.registerPath({
+    method: "get",
+    path: "/org/{orgId}/ai-budgets",
+    description: "List AI budgets for an organization.",
+    tags: [OpenAPITags.AiBudget],
+    request: {
+        params: paramsSchema,
+        query: listSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function listAiBudgets(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedQuery = listSchema.safeParse(req.query);
+        if (!parsedQuery.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedQuery.error).toString()
+                )
+            );
+        }
+
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { orgId } = parsedParams.data;
+        const { pageSize, page } = parsedQuery.data;
+
+        const baseQuery = db
+            .select()
+            .from(aiBudgets)
+            .where(eq(aiBudgets.orgId, orgId));
+
+        const countQuery = db.$count(
+            db
+                .select()
+                .from(aiBudgets)
+                .where(eq(aiBudgets.orgId, orgId))
+                .as("filtered_ai_budgets")
+        );
+
+        const [totalCount, rows] = await Promise.all([
+            countQuery,
+            baseQuery
+                .limit(pageSize)
+                .offset(pageSize * (page - 1))
+                .orderBy(asc(aiBudgets.budgetId))
+        ]);
+
+        return response<ListAiBudgetsResponse>(res, {
+            data: {
+                budgets: rows,
+                pagination: {
+                    total: totalCount,
+                    pageSize,
+                    page
+                }
+            },
+            success: true,
+            error: false,
+            message: "AI budgets retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiBudget/listAiBudgetsForModel.ts

@@ -0,0 +1,69 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiBudgets, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { asc, eq } from "drizzle-orm";
+import type { ListAiBudgetsByScopeResponse } from "@server/routers/aiBudget/types";
+
+const paramsSchema = z.strictObject({
+    modelId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "get",
+    path: "/ai-model/{modelId}/ai-budgets",
+    description: "List AI budgets scoped to an AI model.",
+    tags: [OpenAPITags.AiBudget],
+    request: {
+        params: paramsSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function listAiBudgetsForModel(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { modelId } = parsedParams.data;
+
+        const budgets = await db
+            .select()
+            .from(aiBudgets)
+            .where(eq(aiBudgets.modelId, modelId))
+            .orderBy(asc(aiBudgets.budgetId));
+
+        return response<ListAiBudgetsByScopeResponse>(res, {
+            data: { budgets },
+            success: true,
+            error: false,
+            message: "AI budgets retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiBudget/listAiBudgetsForProvider.ts

@@ -0,0 +1,69 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiBudgets, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { asc, eq } from "drizzle-orm";
+import type { ListAiBudgetsByScopeResponse } from "@server/routers/aiBudget/types";
+
+const paramsSchema = z.strictObject({
+    providerId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "get",
+    path: "/ai-provider/{providerId}/ai-budgets",
+    description: "List AI budgets scoped to an AI provider.",
+    tags: [OpenAPITags.AiBudget],
+    request: {
+        params: paramsSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function listAiBudgetsForProvider(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { providerId } = parsedParams.data;
+
+        const budgets = await db
+            .select()
+            .from(aiBudgets)
+            .where(eq(aiBudgets.providerId, providerId))
+            .orderBy(asc(aiBudgets.budgetId));
+
+        return response<ListAiBudgetsByScopeResponse>(res, {
+            data: { budgets },
+            success: true,
+            error: false,
+            message: "AI budgets retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiBudget/listAiBudgetsForResource.ts

@@ -0,0 +1,69 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiBudgets, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { asc, eq } from "drizzle-orm";
+import type { ListAiBudgetsByScopeResponse } from "@server/routers/aiBudget/types";
+
+const paramsSchema = z.strictObject({
+    resourceId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "get",
+    path: "/resource/{resourceId}/ai-budgets",
+    description: "List AI budgets scoped to a resource.",
+    tags: [OpenAPITags.AiBudget],
+    request: {
+        params: paramsSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function listAiBudgetsForResource(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { resourceId } = parsedParams.data;
+
+        const budgets = await db
+            .select()
+            .from(aiBudgets)
+            .where(eq(aiBudgets.resourceId, resourceId))
+            .orderBy(asc(aiBudgets.budgetId));
+
+        return response<ListAiBudgetsByScopeResponse>(res, {
+            data: { budgets },
+            success: true,
+            error: false,
+            message: "AI budgets retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiBudget/listAiBudgetsForRole.ts

@@ -0,0 +1,69 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiBudgets, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { asc, eq } from "drizzle-orm";
+import type { ListAiBudgetsByScopeResponse } from "@server/routers/aiBudget/types";
+
+const paramsSchema = z.strictObject({
+    roleId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "get",
+    path: "/role/{roleId}/ai-budgets",
+    description: "List AI budgets scoped to a role.",
+    tags: [OpenAPITags.AiBudget],
+    request: {
+        params: paramsSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function listAiBudgetsForRole(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { roleId } = parsedParams.data;
+
+        const budgets = await db
+            .select()
+            .from(aiBudgets)
+            .where(eq(aiBudgets.roleId, roleId))
+            .orderBy(asc(aiBudgets.budgetId));
+
+        return response<ListAiBudgetsByScopeResponse>(res, {
+            data: { budgets },
+            success: true,
+            error: false,
+            message: "AI budgets retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiBudget/listAiBudgetsForSiteResource.ts

@@ -0,0 +1,69 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiBudgets, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { asc, eq } from "drizzle-orm";
+import type { ListAiBudgetsByScopeResponse } from "@server/routers/aiBudget/types";
+
+const paramsSchema = z.strictObject({
+    siteResourceId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "get",
+    path: "/site-resource/{siteResourceId}/ai-budgets",
+    description: "List AI budgets scoped to a site resource.",
+    tags: [OpenAPITags.AiBudget],
+    request: {
+        params: paramsSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function listAiBudgetsForSiteResource(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { siteResourceId } = parsedParams.data;
+
+        const budgets = await db
+            .select()
+            .from(aiBudgets)
+            .where(eq(aiBudgets.siteResourceId, siteResourceId))
+            .orderBy(asc(aiBudgets.budgetId));
+
+        return response<ListAiBudgetsByScopeResponse>(res, {
+            data: { budgets },
+            success: true,
+            error: false,
+            message: "AI budgets retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiBudget/types.ts

@@ -0,0 +1,18 @@
+import type { AiBudget } from "@server/db";
+import type { PaginatedResponse } from "@server/types/Pagination";
+
+export type ListAiBudgetsResponse = PaginatedResponse<{
+    budgets: AiBudget[];
+}>;
+
+export type ListAiBudgetsByScopeResponse = {
+    budgets: AiBudget[];
+};
+
+export type GetAiBudgetResponse = {
+    budget: AiBudget;
+};
+
+export type CreateOrEditAiBudgetResponse = {
+    budget: AiBudget;
+};

server/routers/aiBudget/updateAiBudget.ts

@@ -0,0 +1,343 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import {
+    aiBudgets,
+    aiModels,
+    aiProviders,
+    db,
+    resources,
+    roles,
+    siteResources
+} from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { and, eq, isNull, ne } from "drizzle-orm";
+import type { CreateOrEditAiBudgetResponse } from "@server/routers/aiBudget/types";
+import {
+    aiBudgetEnforcementSchema,
+    aiBudgetPeriodSchema,
+    aiBudgetUnitSchema,
+    refineBudgetScopeFields
+} from "@server/routers/aiBudget/validation";
+
+const paramsSchema = z.strictObject({
+    budgetId: z.coerce.number().int().positive()
+});
+
+const bodySchema = z.strictObject({
+    providerId: z.coerce.number().int().positive().nullable().optional(),
+    modelId: z.coerce.number().int().positive().nullable().optional(),
+    resourceId: z.coerce.number().int().positive().nullable().optional(),
+    siteResourceId: z.coerce.number().int().positive().nullable().optional(),
+    roleId: z.coerce.number().int().positive().nullable().optional(),
+    amount: z.number().positive().optional(),
+    unit: aiBudgetUnitSchema.optional(),
+    period: aiBudgetPeriodSchema.optional(),
+    enforcement: aiBudgetEnforcementSchema.optional(),
+    enabled: z.boolean().optional()
+});
+
+registry.registerPath({
+    method: "post",
+    path: "/ai-budget/{budgetId}",
+    description: "Update an AI budget.",
+    tags: [OpenAPITags.AiBudget],
+    request: {
+        params: paramsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: bodySchema
+                }
+            }
+        }
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function updateAiBudget(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const parsedBody = bodySchema.safeParse(req.body);
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { budgetId } = parsedParams.data;
+        const body = parsedBody.data;
+
+        const [existing] =
+            req.aiBudget && req.aiBudget.budgetId === budgetId
+                ? [req.aiBudget]
+                : await db
+                      .select()
+                      .from(aiBudgets)
+                      .where(eq(aiBudgets.budgetId, budgetId))
+                      .limit(1);
+
+        if (!existing) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI budget with ID ${budgetId} not found`
+                )
+            );
+        }
+
+        const orgId = existing.orgId;
+
+        const nextProviderId =
+            body.providerId !== undefined
+                ? body.providerId
+                : existing.providerId;
+        const nextModelId =
+            body.modelId !== undefined ? body.modelId : existing.modelId;
+        const nextResourceId =
+            body.resourceId !== undefined
+                ? body.resourceId
+                : existing.resourceId;
+        const nextSiteResourceId =
+            body.siteResourceId !== undefined
+                ? body.siteResourceId
+                : existing.siteResourceId;
+        const nextRoleId =
+            body.roleId !== undefined ? body.roleId : existing.roleId;
+        const nextUnit = body.unit !== undefined ? body.unit : existing.unit;
+        const nextPeriod =
+            body.period !== undefined ? body.period : existing.period;
+
+        const scopeValidation = z
+            .object({
+                providerId: z.number().nullable().optional(),
+                modelId: z.number().nullable().optional(),
+                resourceId: z.number().nullable().optional(),
+                siteResourceId: z.number().nullable().optional(),
+                roleId: z.number().nullable().optional()
+            })
+            .superRefine((data, ctx) => refineBudgetScopeFields(data, ctx))
+            .safeParse({
+                providerId: nextProviderId,
+                modelId: nextModelId,
+                resourceId: nextResourceId,
+                siteResourceId: nextSiteResourceId,
+                roleId: nextRoleId
+            });
+
+        if (!scopeValidation.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(scopeValidation.error).toString()
+                )
+            );
+        }
+
+        if (body.providerId !== undefined && body.providerId !== null) {
+            const [provider] = await db
+                .select({ orgId: aiProviders.orgId })
+                .from(aiProviders)
+                .where(eq(aiProviders.providerId, body.providerId))
+                .limit(1);
+            if (!provider || provider.orgId !== orgId) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `AI provider with ID ${body.providerId} not found in this organization`
+                    )
+                );
+            }
+        }
+
+        if (body.modelId !== undefined && body.modelId !== null) {
+            const [model] = await db
+                .select({ orgId: aiProviders.orgId })
+                .from(aiModels)
+                .innerJoin(
+                    aiProviders,
+                    eq(aiModels.providerId, aiProviders.providerId)
+                )
+                .where(eq(aiModels.modelId, body.modelId))
+                .limit(1);
+            if (!model || model.orgId !== orgId) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `AI model with ID ${body.modelId} not found in this organization`
+                    )
+                );
+            }
+        }
+
+        if (body.resourceId !== undefined && body.resourceId !== null) {
+            const [resource] = await db
+                .select({ orgId: resources.orgId })
+                .from(resources)
+                .where(eq(resources.resourceId, body.resourceId))
+                .limit(1);
+            if (!resource || resource.orgId !== orgId) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `Resource with ID ${body.resourceId} not found in this organization`
+                    )
+                );
+            }
+        }
+
+        if (
+            body.siteResourceId !== undefined &&
+            body.siteResourceId !== null
+        ) {
+            const [siteResource] = await db
+                .select({ orgId: siteResources.orgId })
+                .from(siteResources)
+                .where(eq(siteResources.siteResourceId, body.siteResourceId))
+                .limit(1);
+            if (!siteResource || siteResource.orgId !== orgId) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `Site resource with ID ${body.siteResourceId} not found in this organization`
+                    )
+                );
+            }
+        }
+
+        if (body.roleId !== undefined && body.roleId !== null) {
+            const [role] = await db
+                .select({ orgId: roles.orgId })
+                .from(roles)
+                .where(eq(roles.roleId, body.roleId))
+                .limit(1);
+            if (!role || role.orgId !== orgId) {
+                return next(
+                    createHttpError(
+                        HttpCode.NOT_FOUND,
+                        `Role with ID ${body.roleId} not found in this organization`
+                    )
+                );
+            }
+        }
+
+        const scopeCondition =
+            nextProviderId !== null
+                ? eq(aiBudgets.providerId, nextProviderId)
+                : nextModelId !== null
+                  ? eq(aiBudgets.modelId, nextModelId)
+                  : nextResourceId !== null
+                    ? eq(aiBudgets.resourceId, nextResourceId)
+                    : nextSiteResourceId !== null
+                      ? eq(aiBudgets.siteResourceId, nextSiteResourceId)
+                      : nextRoleId !== null
+                        ? eq(aiBudgets.roleId, nextRoleId)
+                        : and(
+                              eq(aiBudgets.orgId, orgId),
+                              isNull(aiBudgets.providerId),
+                              isNull(aiBudgets.modelId),
+                              isNull(aiBudgets.resourceId),
+                              isNull(aiBudgets.siteResourceId),
+                              isNull(aiBudgets.roleId)
+                          );
+
+        const [conflict] = await db
+            .select({ budgetId: aiBudgets.budgetId })
+            .from(aiBudgets)
+            .where(
+                and(
+                    scopeCondition,
+                    eq(aiBudgets.unit, nextUnit),
+                    eq(aiBudgets.period, nextPeriod),
+                    ne(aiBudgets.budgetId, budgetId)
+                )
+            )
+            .limit(1);
+        if (conflict) {
+            return next(
+                createHttpError(
+                    HttpCode.CONFLICT,
+                    `A ${nextPeriod} ${nextUnit} budget already exists for this scope`
+                )
+            );
+        }
+
+        const updateData: Partial<typeof aiBudgets.$inferInsert> = {
+            updatedAt: Date.now()
+        };
+
+        if (body.providerId !== undefined) {
+            updateData.providerId = body.providerId;
+        }
+        if (body.modelId !== undefined) {
+            updateData.modelId = body.modelId;
+        }
+        if (body.resourceId !== undefined) {
+            updateData.resourceId = body.resourceId;
+        }
+        if (body.siteResourceId !== undefined) {
+            updateData.siteResourceId = body.siteResourceId;
+        }
+        if (body.roleId !== undefined) {
+            updateData.roleId = body.roleId;
+        }
+        if (body.amount !== undefined) {
+            updateData.amount = body.amount;
+        }
+        if (body.unit !== undefined) {
+            updateData.unit = body.unit;
+        }
+        if (body.period !== undefined) {
+            updateData.period = body.period;
+        }
+        if (body.enforcement !== undefined) {
+            updateData.enforcement = body.enforcement;
+        }
+        if (body.enabled !== undefined) {
+            updateData.enabled = body.enabled;
+        }
+
+        const [budget] = await db
+            .update(aiBudgets)
+            .set(updateData)
+            .where(eq(aiBudgets.budgetId, budgetId))
+            .returning();
+
+        return response<CreateOrEditAiBudgetResponse>(res, {
+            data: { budget },
+            success: true,
+            error: false,
+            message: "AI budget updated successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiBudget/validation.ts

@@ -0,0 +1,46 @@
+import { z } from "zod";
+
+export const aiBudgetUnitSchema = z.enum(["usd", "tokens"]);
+
+export const aiBudgetPeriodSchema = z.enum([
+    "monthly",
+    "yearly",
+    "lifetime",
+    "daily",
+    "hourly",
+    "weekly"
+]);
+
+export const aiBudgetEnforcementSchema = z.enum(["hard", "soft"]);
+
+export function refineBudgetScopeFields(
+    data: {
+        providerId?: number | null;
+        modelId?: number | null;
+        resourceId?: number | null;
+        siteResourceId?: number | null;
+        roleId?: number | null;
+    },
+    ctx: z.RefinementCtx
+) {
+    const scopeFields = [
+        data.providerId,
+        data.modelId,
+        data.resourceId,
+        data.siteResourceId,
+        data.roleId
+    ];
+
+    const setCount = scopeFields.filter(
+        (value) => value !== null && value !== undefined
+    ).length;
+
+    if (setCount > 1) {
+        ctx.addIssue({
+            code: "custom",
+            message:
+                "Only one of providerId, modelId, resourceId, siteResourceId, or roleId may be set on a budget",
+            path: ["providerId"]
+        });
+    }
+}

server/routers/aiGateway/createAiGatewayRouter.ts

@@ -0,0 +1,21 @@
+import { Router } from "express";
+import {
+    AI_CAPABILITY_DEFS,
+    type AiCapability
+} from "@server/lib/aiCapabilities";
+import { handleAiGatewayProxy } from "@server/routers/aiGateway/pipeline";
+
+export function createAiGatewayRouter() {
+    const router = Router();
+
+    for (const def of Object.values(AI_CAPABILITY_DEFS)) {
+        const capability = def.id as AiCapability;
+        for (const route of def.routes) {
+            router.post(route.path, (req, res) =>
+                handleAiGatewayProxy(req, res, capability)
+            );
+        }
+    }
+
+    return router;
+}

server/routers/aiGateway/index.ts

@@ -0,0 +1,2 @@
+export { handleAiGatewayProxy } from "./pipeline";
+export { createAiGatewayRouter } from "./createAiGatewayRouter";

server/routers/aiGateway/pipeline.ts

@@ -0,0 +1,841 @@
+import { Request, Response } from "express";
+import { and, eq, inArray } from "drizzle-orm";
+import {
+    AiProvider,
+    aiModels,
+    aiProviders,
+    clients,
+    db,
+    exitNodes,
+    resourceAiModels,
+    resourceAiProviders,
+    resources,
+    siteResourceAiModels,
+    siteResourceAiProviders,
+    siteResources,
+    users
+} from "@server/db";
+import config from "@server/lib/config";
+import { decrypt } from "@server/lib/crypto";
+import {
+    AiProviderAuthType,
+    AiProviderType,
+    applyAiProviderAuthHeaders,
+    applyAiProviderCustomHeaders,
+    authTypeRequiresApiKey
+} from "@server/lib/aiProviderDefaults";
+import {
+    AI_CAPABILITY_DEFS,
+    providerHasCapability,
+    type AiCapability
+} from "@server/lib/aiCapabilities";
+import { proxyAiGatewayToSiteTarget } from "@server/routers/aiGateway/targetRouting";
+import {
+    SESSION_COOKIE_NAME,
+    validateSessionToken
+} from "@server/auth/sessions/app";
+import { getUserOrgRoles } from "@server/lib/userOrgRoles";
+import { isIpInCidr } from "@server/lib/ip";
+import { localCache } from "@server/lib/cache";
+import logger from "@server/logger";
+import HttpCode from "@server/types/HttpCode";
+import {
+    resolveEffectiveLists,
+    type AccessMode,
+    type ModelListType
+} from "@server/lib/aiInferenceResource";
+import {
+    compareModelKeySpecificity,
+    isAllowedByLists,
+    mostSpecificMatchingAllow
+} from "@server/lib/aiModelKeyMatch";
+import { aiGatewayUpstreamFetch } from "@server/lib/aiGatewayUpstreamFetch";
+import { getModelPricing, calculateAiCost } from "@server/lib/aiModelPricing";
+import {
+    extractUsage,
+    estimateUsage,
+    isUsageEmpty,
+    needsStreamUsageInjection,
+    withStreamUsageOption,
+    stripInjectedUsageFrame,
+    extractResponseModel,
+    type AiUsage
+} from "@server/lib/aiUsageExtraction";
+
+const EXIT_NODE_RANGES_CACHE_KEY = "aiGateway:exitNodeRanges";
+const EXIT_NODE_RANGES_TTL_SEC = 6000;
+const CLIENT_BY_IP_TTL_SEC = 30;
+const REQUEST_USER_TTL_SEC = 30;
+
+type CachedClient = { clientId: number; userId: string | null } | null;
+
+async function getExitNodeRanges(): Promise<string[]> {
+    const cached = localCache.get<string[]>(EXIT_NODE_RANGES_CACHE_KEY);
+    if (cached) {
+        return cached;
+    }
+
+    const rows = await db
+        .select({ address: exitNodes.address })
+        .from(exitNodes);
+    const ranges = rows.map((r) => r.address);
+
+    localCache.set(
+        EXIT_NODE_RANGES_CACHE_KEY,
+        ranges,
+        EXIT_NODE_RANGES_TTL_SEC
+    );
+    return ranges;
+}
+
+async function findClientByIp(ip: string): Promise<CachedClient> {
+    const cacheKey = `aiGateway:clientByIp:${ip}`;
+    const cached = localCache.get<CachedClient>(cacheKey);
+    if (cached !== undefined) {
+        return cached;
+    }
+
+    const [client] = await db
+        .select({ clientId: clients.clientId, userId: clients.userId })
+        .from(clients)
+        .where(eq(clients.exitNodeSubnet, `${ip}/32`))
+        .limit(1);
+
+    const result: CachedClient = client || null;
+    localCache.set(cacheKey, result, CLIENT_BY_IP_TTL_SEC);
+    return result;
+}
+
+type ProviderAttachment = {
+    provider: AiProvider;
+    accessMode: AccessMode;
+};
+
+type ResourceModelPattern = {
+    providerId: number;
+    modelKey: string;
+    listType: ModelListType;
+    enabled: boolean;
+};
+
+type ProviderPatternLists = {
+    allows: string[];
+    blocks: string[];
+};
+
+type ResolvedTarget = {
+    resourceId: number | null;
+    siteResourceId: number | null;
+    orgId: string | null;
+    attachments: ProviderAttachment[];
+    resourceListsByProvider: Map<number, ProviderPatternLists>;
+};
+
+type ProviderSelection =
+    | { ok: true; provider: AiProvider }
+    | { ok: false; status: number; message: string };
+
+export type RequestUser = {
+    userId: string;
+    username: string;
+    email: string | null;
+    name: string | null;
+    role: string | null;
+};
+
+// Identity headers forwarded to the upstream inference endpoint when the
+// requesting user is known. Omitted entirely (not sent empty) when we
+// couldn't resolve a user for the request.
+export function applyRequestUserHeaders(
+    headers: Record<string, string>,
+    requestUser: RequestUser | null
+): void {
+    if (!requestUser) {
+        return;
+    }
+    headers["Remote-User"] = requestUser.username;
+    if (requestUser.email) {
+        headers["Remote-Email"] = requestUser.email;
+    }
+    if (requestUser.name) {
+        headers["Remote-Name"] = requestUser.name;
+    }
+    if (requestUser.role) {
+        headers["Remote-Role"] = requestUser.role;
+    }
+}
+
+async function buildRequestUser(
+    userId: string,
+    orgId: string | null
+): Promise<RequestUser | null> {
+    const cacheKey = `aiGateway:requestUser:${userId}:${orgId || ""}`;
+    const cached = localCache.get<RequestUser | null>(cacheKey);
+    if (cached !== undefined) {
+        return cached;
+    }
+
+    const [user] = await db
+        .select()
+        .from(users)
+        .where(eq(users.userId, userId))
+        .limit(1);
+
+    if (!user) {
+        localCache.set(cacheKey, null, REQUEST_USER_TTL_SEC);
+        return null;
+    }
+
+    const orgRoles = orgId ? await getUserOrgRoles(user.userId, orgId) : [];
+
+    const requestUser: RequestUser = {
+        userId: user.userId,
+        username: user.username,
+        email: user.email,
+        name: user.name,
+        role: orgRoles.map((r) => r.roleName).join(", ") || null
+    };
+
+    localCache.set(cacheKey, requestUser, REQUEST_USER_TTL_SEC);
+    return requestUser;
+}
+
+async function resolveRequestUser(
+    req: Request,
+    _resourceId: number | null,
+    orgId: string | null
+): Promise<RequestUser | null> {
+    const sessionToken = req.cookies?.[SESSION_COOKIE_NAME];
+    if (sessionToken) {
+        const { session, user } = await validateSessionToken(sessionToken);
+        if (session && user) {
+            return buildRequestUser(user.userId, orgId);
+        }
+    }
+
+    // TODO: MAKE SURE THIS CAN NOT BE SPOOFED AND CAN BE TRUSTED AS AN INTERNAL ADDRESS FROM A NODE
+
+    const ip = req.ip;
+    if (!ip) {
+        return null;
+    }
+
+    const exitNodeRanges = await getExitNodeRanges();
+    const inExitNodeRange = exitNodeRanges.some((range) =>
+        isIpInCidr(ip, range)
+    );
+    if (!inExitNodeRange) {
+        return null;
+    }
+
+    const client = await findClientByIp(ip);
+    if (!client || !client.userId) {
+        return null;
+    }
+
+    return buildRequestUser(client.userId, orgId);
+}
+
+async function resolveTarget(host: string): Promise<ResolvedTarget | null> {
+    const [[resourceRow], [siteResourceRow]] = await Promise.all([
+        db
+            .select({
+                resourceId: resources.resourceId,
+                orgId: resources.orgId
+            })
+            .from(resources)
+            .where(
+                and(
+                    eq(resources.fullDomain, host),
+                    eq(resources.mode, "inference"),
+                    eq(resources.enabled, true)
+                )
+            )
+            .limit(1),
+        db
+            .select({
+                siteResourceId: siteResources.siteResourceId,
+                orgId: siteResources.orgId
+            })
+            .from(siteResources)
+            .where(
+                and(
+                    eq(siteResources.fullDomain, host),
+                    eq(siteResources.mode, "inference"),
+                    eq(siteResources.enabled, true)
+                )
+            )
+            .limit(1)
+    ]);
+
+    // Prefer public inference resources when both match the same host.
+    if (resourceRow) {
+        const [attachmentRows, resourcePatterns] = await Promise.all([
+            db
+                .select({
+                    provider: aiProviders,
+                    accessMode: resourceAiProviders.accessMode
+                })
+                .from(resourceAiProviders)
+                .innerJoin(
+                    aiProviders,
+                    eq(resourceAiProviders.providerId, aiProviders.providerId)
+                )
+                .where(
+                    and(
+                        eq(
+                            resourceAiProviders.resourceId,
+                            resourceRow.resourceId
+                        ),
+                        eq(aiProviders.enabled, true),
+                        eq(resourceAiProviders.enabled, true)
+                    )
+                ),
+            db
+                .select({
+                    providerId: aiModels.providerId,
+                    modelKey: aiModels.modelKey,
+                    listType: resourceAiModels.listType,
+                    enabled: aiModels.enabled
+                })
+                .from(resourceAiModels)
+                .innerJoin(
+                    aiModels,
+                    eq(resourceAiModels.modelId, aiModels.modelId)
+                )
+                .where(eq(resourceAiModels.resourceId, resourceRow.resourceId))
+        ]);
+
+        if (attachmentRows.length === 0) {
+            return null;
+        }
+
+        return {
+            resourceId: resourceRow.resourceId,
+            siteResourceId: null,
+            orgId: resourceRow.orgId,
+            attachments: attachmentRows.map((a) => ({
+                provider: a.provider,
+                accessMode: a.accessMode
+            })),
+            resourceListsByProvider: groupPatternsByProvider(resourcePatterns)
+        };
+    }
+
+    if (siteResourceRow) {
+        const [attachmentRows, resourcePatterns] = await Promise.all([
+            db
+                .select({
+                    provider: aiProviders,
+                    accessMode: siteResourceAiProviders.accessMode
+                })
+                .from(siteResourceAiProviders)
+                .innerJoin(
+                    aiProviders,
+                    eq(
+                        siteResourceAiProviders.providerId,
+                        aiProviders.providerId
+                    )
+                )
+                .where(
+                    and(
+                        eq(
+                            siteResourceAiProviders.siteResourceId,
+                            siteResourceRow.siteResourceId
+                        ),
+                        eq(aiProviders.enabled, true),
+                        eq(siteResourceAiProviders.enabled, true)
+                    )
+                ),
+            db
+                .select({
+                    providerId: aiModels.providerId,
+                    modelKey: aiModels.modelKey,
+                    listType: siteResourceAiModels.listType,
+                    enabled: aiModels.enabled
+                })
+                .from(siteResourceAiModels)
+                .innerJoin(
+                    aiModels,
+                    eq(siteResourceAiModels.modelId, aiModels.modelId)
+                )
+                .where(
+                    eq(
+                        siteResourceAiModels.siteResourceId,
+                        siteResourceRow.siteResourceId
+                    )
+                )
+        ]);
+
+        if (attachmentRows.length === 0) {
+            return null;
+        }
+
+        return {
+            resourceId: null,
+            siteResourceId: siteResourceRow.siteResourceId,
+            orgId: siteResourceRow.orgId,
+            attachments: attachmentRows.map((a) => ({
+                provider: a.provider,
+                accessMode: a.accessMode
+            })),
+            resourceListsByProvider: groupPatternsByProvider(resourcePatterns)
+        };
+    }
+
+    return null;
+}
+
+function groupPatternsByProvider(
+    patterns: ResourceModelPattern[]
+): Map<number, ProviderPatternLists> {
+    const byProvider = new Map<number, ProviderPatternLists>();
+    for (const pattern of patterns) {
+        if (!pattern.enabled) {
+            continue;
+        }
+        let lists = byProvider.get(pattern.providerId);
+        if (!lists) {
+            lists = { allows: [], blocks: [] };
+            byProvider.set(pattern.providerId, lists);
+        }
+        if (pattern.listType === "allow") {
+            lists.allows.push(pattern.modelKey);
+        } else {
+            lists.blocks.push(pattern.modelKey);
+        }
+    }
+    return byProvider;
+}
+
+async function selectProvider(
+    attachments: ProviderAttachment[],
+    resourceListsByProvider: Map<number, ProviderPatternLists>,
+    requestedModel: string | undefined
+): Promise<ProviderSelection> {
+    if (!requestedModel) {
+        return {
+            ok: false,
+            status: HttpCode.FORBIDDEN,
+            message: "A model must be specified for this resource"
+        };
+    }
+
+    const attachmentByProviderId = new Map(
+        attachments.map((a) => [a.provider.providerId, a])
+    );
+    const providerIds = [...attachmentByProviderId.keys()];
+    if (providerIds.length === 0) {
+        return {
+            ok: false,
+            status: HttpCode.FORBIDDEN,
+            message: `Model "${requestedModel}" is not permitted on this resource`
+        };
+    }
+
+    const providerModels = await db
+        .select({
+            providerId: aiModels.providerId,
+            modelKey: aiModels.modelKey,
+            listType: aiModels.listType,
+            enabled: aiModels.enabled
+        })
+        .from(aiModels)
+        .where(inArray(aiModels.providerId, providerIds));
+
+    const allowsByProvider = new Map<number, string[]>();
+    const blocksByProvider = new Map<number, string[]>();
+    for (const model of providerModels) {
+        if (!model.enabled) {
+            continue;
+        }
+        const targetMap =
+            model.listType === "allow" ? allowsByProvider : blocksByProvider;
+        const existing = targetMap.get(model.providerId) ?? [];
+        existing.push(model.modelKey);
+        targetMap.set(model.providerId, existing);
+    }
+
+    type ModelCandidate = {
+        provider: AiProvider;
+        modelKey: string;
+    };
+
+    const candidates: ModelCandidate[] = [];
+    for (const [providerId, attachment] of attachmentByProviderId) {
+        const resourceLists = resourceListsByProvider.get(providerId);
+        const { allows, blocks } = resolveEffectiveLists({
+            accessMode: attachment.accessMode,
+            providerAllows: allowsByProvider.get(providerId) ?? [],
+            providerBlocks: blocksByProvider.get(providerId) ?? [],
+            resourceAllows: resourceLists?.allows ?? [],
+            resourceBlocks: resourceLists?.blocks ?? []
+        });
+
+        if (!isAllowedByLists(requestedModel, allows, blocks)) {
+            continue;
+        }
+        const matchingAllow = mostSpecificMatchingAllow(requestedModel, allows);
+        if (!matchingAllow) {
+            continue;
+        }
+        candidates.push({
+            provider: attachment.provider,
+            modelKey: matchingAllow
+        });
+    }
+
+    if (candidates.length === 0) {
+        return {
+            ok: false,
+            status: HttpCode.FORBIDDEN,
+            message: `Model "${requestedModel}" is not permitted on this resource`
+        };
+    }
+
+    candidates.sort((a, b) =>
+        compareModelKeySpecificity(a.modelKey, b.modelKey)
+    );
+
+    const bestSpecificity = candidates[0].modelKey;
+    const topCandidates = candidates.filter(
+        (c) => compareModelKeySpecificity(c.modelKey, bestSpecificity) === 0
+    );
+
+    const uniqueProviders = new Map<number, AiProvider>();
+    for (const candidate of topCandidates) {
+        uniqueProviders.set(candidate.provider.providerId, candidate.provider);
+    }
+
+    if (uniqueProviders.size === 1) {
+        return { ok: true, provider: [...uniqueProviders.values()][0] };
+    }
+
+    return {
+        ok: false,
+        status: HttpCode.FORBIDDEN,
+        message: `Model "${requestedModel}" is ambiguous across multiple AI providers on this resource`
+    };
+}
+
+function logAiUsageAndCost(args: {
+    capability: AiCapability;
+    provider: AiProvider;
+    requestedModel: string | undefined;
+    requestBody: unknown;
+    responseText: string;
+    isStream: boolean;
+    headers: Headers;
+}): void {
+    const {
+        capability,
+        provider,
+        requestedModel,
+        requestBody,
+        responseText,
+        isStream,
+        headers
+    } = args;
+
+    let usage: AiUsage | null = extractUsage(
+        capability,
+        responseText,
+        isStream,
+        headers
+    );
+    if (!usage || isUsageEmpty(usage)) {
+        usage = estimateUsage(JSON.stringify(requestBody ?? ""), responseText);
+    }
+
+    const model = extractResponseModel(responseText) ?? requestedModel;
+    const pricing = getModelPricing(provider.type as AiProviderType, model);
+    const cost = calculateAiCost(pricing, usage);
+
+    logger.info("AI gateway request usage", {
+        capability,
+        providerId: provider.providerId,
+        providerType: provider.type,
+        model,
+        estimated: usage.estimated,
+        promptTokens: usage.promptTokens,
+        cacheReadTokens: usage.cacheReadTokens,
+        cacheWriteTokens: usage.cacheWriteTokens,
+        completionTokens: usage.completionTokens,
+        reasoningTokens: usage.reasoningTokens,
+        pricingApproximate: pricing?.approximate ?? null,
+        totalCostUsd: cost?.totalCost ?? null
+    });
+}
+
+export async function handleAiGatewayProxy(
+    req: Request,
+    res: Response,
+    capability: AiCapability
+): Promise<any> {
+    try {
+        const def = AI_CAPABILITY_DEFS[capability];
+
+        const host = (
+            (req.headers["p-host"] as string | undefined) || // p-host is only used sometimes when overriding the host header for some middleware proxy
+            req.headers.host ||
+            ""
+        ).split(":")[0];
+        if (!host) {
+            return res
+                .status(HttpCode.BAD_REQUEST)
+                .json({ error: { message: "Missing Host header" } });
+        }
+
+        logger.info(`AI gateway ${capability} request for host: ${host}`);
+
+        const target = await resolveTarget(host);
+        if (!target) {
+            return res.status(HttpCode.NOT_FOUND).json({
+                error: {
+                    message: "No inference resource found for this host"
+                }
+            });
+        }
+
+        const { attachments, resourceListsByProvider, resourceId, orgId } =
+            target;
+
+        const capableAttachments = attachments.filter((a) =>
+            providerHasCapability(a.provider.capabilities, capability)
+        );
+
+        if (capableAttachments.length === 0) {
+            return res.status(HttpCode.FORBIDDEN).json({
+                error: {
+                    message: `No AI provider on this resource supports ${capability}`
+                }
+            });
+        }
+
+        const requestedModel = def.extractModel(req);
+
+        const [requestUser, selection] = await Promise.all([
+            resolveRequestUser(req, resourceId, orgId),
+            selectProvider(
+                capableAttachments,
+                resourceListsByProvider,
+                requestedModel
+            )
+        ]);
+
+        if (requestUser) {
+            logger.debug(
+                `AI gateway request from user ${requestUser.userId} (${requestUser.username})`
+            );
+        }
+
+        if (!selection.ok) {
+            return res.status(selection.status).json({
+                error: { message: selection.message }
+            });
+        }
+
+        const { provider } = selection;
+
+        if (provider.type === "custom" && provider.routingMode === "target") {
+            return await proxyAiGatewayToSiteTarget(
+                req,
+                res,
+                provider,
+                requestUser,
+                capability
+            );
+        }
+
+        const upstreamUrl = provider.upstreamUrl;
+        const authType = provider.authType as AiProviderAuthType;
+
+        if (!upstreamUrl) {
+            return res.status(HttpCode.INTERNAL_SERVER_ERROR).json({
+                error: {
+                    message: "AI provider has no upstream URL configured"
+                }
+            });
+        }
+
+        let apiKey: string | null = null;
+        if (authTypeRequiresApiKey(authType)) {
+            if (!provider.apiKey) {
+                return res.status(HttpCode.INTERNAL_SERVER_ERROR).json({
+                    error: {
+                        message: "AI provider has no API key configured"
+                    }
+                });
+            }
+            const secret = config.getRawConfig().server.secret!;
+            apiKey = decrypt(provider.apiKey, secret);
+        }
+
+        const targetUrl = def.resolveUpstreamUrl(
+            upstreamUrl,
+            req,
+            requestedModel!
+        );
+
+        const skipHeaders = new Set([
+            "p-host",
+            "host",
+            "connection",
+            "keep-alive",
+            "proxy-authenticate",
+            "proxy-authorization",
+            "te",
+            "trailers",
+            "transfer-encoding",
+            "upgrade",
+            "content-length",
+            "accept-encoding"
+        ]);
+
+        const headers: Record<string, string> = {};
+        for (const [key, value] of Object.entries(req.headers)) {
+            if (skipHeaders.has(key.toLowerCase()) || value === undefined) {
+                continue;
+            }
+            headers[key] = Array.isArray(value) ? value.join(", ") : value;
+        }
+        applyAiProviderCustomHeaders(
+            headers,
+            provider.headers,
+            config.getRawConfig().server.secret!
+        );
+        applyAiProviderAuthHeaders(headers, authType, apiKey);
+        applyRequestUserHeaders(headers, requestUser);
+
+        // OpenAI's Chat Completions API only reports usage in a streaming
+        // response when asked to via stream_options.include_usage - inject
+        // it ourselves when the caller didn't, so we can still track cost,
+        // and strip the extra frame it adds back out of what we forward.
+        const injectedUsageOurselves = needsStreamUsageInjection(
+            capability,
+            req.body
+        );
+        const outboundBody = injectedUsageOurselves
+            ? withStreamUsageOption(req.body)
+            : req.body;
+        const body = JSON.stringify(outboundBody);
+
+        logger.debug("AI gateway upstream request", {
+            capability,
+            url: targetUrl,
+            method: "POST",
+            headers,
+            body: outboundBody,
+            skipTlsVerification: provider.skipTlsVerification
+        });
+
+        const abortController = new AbortController();
+        const onClientClose = () => {
+            if (!res.writableEnded) {
+                abortController.abort();
+            }
+        };
+        res.on("close", onClientClose);
+
+        let upstreamRes: globalThis.Response;
+        try {
+            upstreamRes = await aiGatewayUpstreamFetch(targetUrl, {
+                method: "POST",
+                headers,
+                body,
+                skipTlsVerification: provider.skipTlsVerification,
+                signal: abortController.signal
+            });
+        } catch (fetchError) {
+            res.off("close", onClientClose);
+            if (abortController.signal.aborted) {
+                // Client already disconnected; nothing left to respond to.
+                return;
+            }
+            logger.error({
+                message: "AI gateway upstream fetch failed",
+                url: targetUrl,
+                error: fetchError,
+                cause:
+                    fetchError instanceof Error
+                        ? (fetchError as Error & { cause?: unknown }).cause
+                        : undefined
+            });
+            throw fetchError;
+        }
+
+        const contentType = upstreamRes.headers.get("content-type") || "";
+        const isStream = def.isStreaming(req, contentType);
+
+        res.status(upstreamRes.status);
+        res.setHeader("Content-Type", contentType || "application/json");
+
+        if (isStream && upstreamRes.body) {
+            res.flushHeaders();
+            const reader = upstreamRes.body.getReader();
+            const decoder = new TextDecoder();
+            let fullText = "";
+            // Frame-boundary buffer, only used when we need to filter the
+            // usage-only frame we injected out of what reaches the client.
+            let sseCarry = "";
+            try {
+                while (!abortController.signal.aborted) {
+                    const { done, value } = await reader.read();
+                    if (done) break;
+                    const chunkText = decoder.decode(value, { stream: true });
+                    fullText += chunkText;
+                    if (injectedUsageOurselves) {
+                        sseCarry += chunkText;
+                        const lastBoundary = sseCarry.lastIndexOf("\n\n");
+                        if (lastBoundary !== -1) {
+                            const toEmit = sseCarry.slice(0, lastBoundary + 2);
+                            sseCarry = sseCarry.slice(lastBoundary + 2);
+                            res.write(stripInjectedUsageFrame(toEmit));
+                        }
+                    } else {
+                        res.write(value);
+                    }
+                }
+                if (injectedUsageOurselves && sseCarry) {
+                    res.write(stripInjectedUsageFrame(sseCarry));
+                }
+            } finally {
+                await reader.cancel().catch(() => {});
+                res.off("close", onClientClose);
+            }
+            if (!res.writableEnded) {
+                res.end();
+            }
+            if (!abortController.signal.aborted) {
+                logAiUsageAndCost({
+                    capability,
+                    provider,
+                    requestedModel,
+                    requestBody: req.body,
+                    responseText: fullText,
+                    isStream: true,
+                    headers: upstreamRes.headers
+                });
+            }
+            return;
+        }
+
+        res.off("close", onClientClose);
+        const text = await upstreamRes.text();
+        logAiUsageAndCost({
+            capability,
+            provider,
+            requestedModel,
+            requestBody: req.body,
+            responseText: text,
+            isStream: false,
+            headers: upstreamRes.headers
+        });
+        return res.send(text);
+    } catch (error) {
+        logger.error(error);
+        return res.status(HttpCode.INTERNAL_SERVER_ERROR).json({
+            error: { message: "Failed to proxy inference request" }
+        });
+    }
+}

server/routers/aiGateway/targetRouting.ts

@@ -0,0 +1,293 @@
+import { Request, Response } from "express";
+import { and, eq } from "drizzle-orm";
+import { AiProvider, db, exitNodes, sites, targets } from "@server/db";
+import config from "@server/lib/config";
+import { decrypt } from "@server/lib/crypto";
+import { localCache } from "@server/lib/cache";
+import {
+    AiProviderAuthType,
+    applyAiProviderAuthHeaders,
+    applyAiProviderCustomHeaders,
+    authTypeRequiresApiKey
+} from "@server/lib/aiProviderDefaults";
+import {
+    AI_CAPABILITY_DEFS,
+    type AiCapability
+} from "@server/lib/aiCapabilities";
+import logger from "@server/logger";
+import HttpCode from "@server/types/HttpCode";
+import {
+    applyRequestUserHeaders,
+    type RequestUser
+} from "@server/routers/aiGateway/pipeline";
+
+// Short TTL: long enough to spare the DB on a burst of requests, short
+// enough that target/site changes (added, removed, exit node moved) show up
+// almost immediately without needing explicit cache invalidation.
+const PROVIDER_TARGETS_TTL_SEC = 7;
+
+// Header gerbil reads to know which scheme://host:port (reachable over the
+// WireGuard network) to rewrite an incoming /router/* request to. Must
+// match gerbil's `pangolinDestHeader` constant.
+const PANGOLIN_DEST_HEADER = "p-dest-header";
+
+// Header gerbil reads for the Host header value to send to the destination,
+// when it should differ from PANGOLIN_DEST_HEADER (the target's configured
+// ip rather than the WireGuard routing address). Must match gerbil's
+// `pangolinHostHeader` constant.
+const PANGOLIN_HOST_HEADER = "p-dest-host-header";
+
+const SKIP_HEADERS = new Set([
+    "p-host",
+    "host",
+    "connection",
+    "keep-alive",
+    "proxy-authenticate",
+    "proxy-authorization",
+    "te",
+    "trailers",
+    "transfer-encoding",
+    "upgrade",
+    "content-length",
+    "accept-encoding"
+]);
+
+type ResolvedProviderTarget = {
+    targetId: number;
+    // "<scheme>://<site exitNodeSubnet host>:<internalPort>", passed to
+    // gerbil as the destination to proxy the request to over the WireGuard
+    // tunnel.
+    destination: string;
+    // The target's configured ip, passed to gerbil as the Host header to
+    // send to the destination (which may differ from the WireGuard routing
+    // address above, e.g. for vhost-based targets).
+    hostHeader: string;
+    // The target's site's exit node HTTP API base URL (gerbil's /router/*).
+    gerbilBaseUrl: string;
+};
+
+async function fetchProviderTargets(
+    providerId: number
+): Promise<ResolvedProviderTarget[]> {
+    const rows = await db
+        .select({
+            targetId: targets.targetId,
+            ip: targets.ip,
+            internalPort: targets.internalPort,
+            port: targets.port,
+            method: targets.method,
+            exitNodeSubnet: sites.exitNodeSubnet,
+            reachableAt: exitNodes.reachableAt
+        })
+        .from(targets)
+        .innerJoin(sites, eq(targets.siteId, sites.siteId))
+        .innerJoin(exitNodes, eq(sites.exitNodeId, exitNodes.exitNodeId))
+        .where(
+            and(eq(targets.providerId, providerId), eq(targets.enabled, true))
+        );
+
+    const resolved: ResolvedProviderTarget[] = [];
+    for (const row of rows) {
+        // Sites not yet connected to an exit node (no subnet assigned) or
+        // whose exit node has no known HTTP address can't be routed to.
+        if (!row.exitNodeSubnet || !row.reachableAt) {
+            continue;
+        }
+        const host = row.exitNodeSubnet.split("/")[0];
+        const port = row.internalPort ?? row.port;
+        const scheme = row.method?.toLowerCase() ?? "https";
+        resolved.push({
+            targetId: row.targetId,
+            destination: `${scheme}://${host}:${port}`,
+            hostHeader: row.ip,
+            gerbilBaseUrl: row.reachableAt
+        });
+    }
+
+    return resolved;
+}
+
+async function getProviderTargets(
+    providerId: number
+): Promise<ResolvedProviderTarget[]> {
+    const cacheKey = `aiGateway:providerTargets:${providerId}`;
+    const cached = localCache.get<ResolvedProviderTarget[]>(cacheKey);
+    if (cached !== undefined) {
+        return cached;
+    }
+
+    const resolved = await fetchProviderTargets(providerId);
+    localCache.set(cacheKey, resolved, PROVIDER_TARGETS_TTL_SEC);
+    return resolved;
+}
+
+// Round-robin cursor per provider. Process-local and unpersisted - fine
+// since it only needs to spread load across targets, not guarantee a
+// perfectly even distribution across restarts or multiple server instances.
+const roundRobinCursors = new Map<number, number>();
+
+function pickTarget(
+    providerId: number,
+    providerTargets: ResolvedProviderTarget[]
+): ResolvedProviderTarget {
+    const cursor = roundRobinCursors.get(providerId) ?? 0;
+    roundRobinCursors.set(providerId, cursor + 1);
+    return providerTargets[cursor % providerTargets.length];
+}
+
+function pathFromRequest(req: Request): string {
+    // Query string is preserved - some providers use it to select the
+    // streaming response format (e.g. Gemini's `?alt=sse`), and gerbil's
+    // /router/* forwards it through untouched.
+    const raw = req.originalUrl || req.url || req.path;
+    return raw.startsWith("/") ? raw : `/${raw}`;
+}
+
+/**
+ * Proxies an AI gateway request to one of a "custom" / "target" routing-mode
+ * provider's site targets, via that site's gerbil sidecar. Gerbil's
+ * /router/* endpoint forwards the request (untouched body, same path minus
+ * the /router prefix, and all headers besides PANGOLIN_DEST_HEADER and
+ * PANGOLIN_HOST_HEADER) over the WireGuard tunnel to the destination named
+ * in PANGOLIN_DEST_HEADER, sending PANGOLIN_HOST_HEADER as the Host header.
+ * Always writes a response to `res`, including on failure.
+ */
+export async function proxyAiGatewayToSiteTarget(
+    req: Request,
+    res: Response,
+    provider: AiProvider,
+    requestUser: RequestUser | null,
+    capability: AiCapability
+): Promise<void> {
+    const providerTargets = await getProviderTargets(provider.providerId);
+    if (providerTargets.length === 0) {
+        res.status(HttpCode.INTERNAL_SERVER_ERROR).json({
+            error: {
+                message: "AI provider has no reachable site targets configured"
+            }
+        });
+        return;
+    }
+
+    const target = pickTarget(provider.providerId, providerTargets);
+    const gerbilUrl = `${target.gerbilBaseUrl.replace(/\/+$/, "")}/router${pathFromRequest(req)}`;
+
+    const headers: Record<string, string> = {};
+    for (const [key, value] of Object.entries(req.headers)) {
+        if (SKIP_HEADERS.has(key.toLowerCase()) || value === undefined) {
+            continue;
+        }
+        headers[key] = Array.isArray(value) ? value.join(", ") : value;
+    }
+
+    const authType = provider.authType as AiProviderAuthType;
+    let apiKey: string | null = null;
+    if (authTypeRequiresApiKey(authType)) {
+        if (!provider.apiKey) {
+            res.status(HttpCode.INTERNAL_SERVER_ERROR).json({
+                error: {
+                    message: "AI provider has no API key configured"
+                }
+            });
+            return;
+        }
+        const secret = config.getRawConfig().server.secret!;
+        apiKey = decrypt(provider.apiKey, secret);
+    }
+    applyAiProviderCustomHeaders(
+        headers,
+        provider.headers,
+        config.getRawConfig().server.secret!
+    );
+    applyAiProviderAuthHeaders(headers, authType, apiKey);
+    applyRequestUserHeaders(headers, requestUser);
+
+    headers[PANGOLIN_DEST_HEADER] = target.destination;
+    headers[PANGOLIN_HOST_HEADER] = target.hostHeader;
+
+    const body = JSON.stringify(req.body);
+
+    logger.debug("AI gateway target-routed request", {
+        providerId: provider.providerId,
+        targetId: target.targetId,
+        destination: target.destination,
+        hostHeader: target.hostHeader,
+        url: gerbilUrl,
+        headers,
+        body: req.body
+    });
+
+    // Cancel the request to gerbil (which cascades to gerbil cancelling its
+    // proxied request to the actual site target, since gerbil's reverse
+    // proxy derives the outbound request's context from the inbound one) if
+    // the client goes away before we're done.
+    const abortController = new AbortController();
+    const onClientClose = () => {
+        if (!res.writableEnded) {
+            abortController.abort();
+        }
+    };
+    res.on("close", onClientClose);
+
+    let upstreamRes: globalThis.Response;
+    try {
+        upstreamRes = await fetch(gerbilUrl, {
+            method: "POST",
+            headers,
+            body,
+            signal: abortController.signal
+        });
+    } catch (fetchError) {
+        res.off("close", onClientClose);
+        if (abortController.signal.aborted) {
+            // Client already disconnected; nothing left to respond to.
+            return;
+        }
+        logger.error({
+            message: "AI gateway target proxy request failed",
+            url: gerbilUrl,
+            targetId: target.targetId,
+            error: fetchError,
+            cause:
+                fetchError instanceof Error
+                    ? (fetchError as Error & { cause?: unknown }).cause
+                    : undefined
+        });
+        res.status(HttpCode.BAD_GATEWAY).json({
+            error: { message: "Failed to reach AI provider target" }
+        });
+        return;
+    }
+
+    const contentType = upstreamRes.headers.get("content-type") || "";
+    const isStream = AI_CAPABILITY_DEFS[capability].isStreaming(
+        req,
+        contentType
+    );
+
+    res.status(upstreamRes.status);
+    res.setHeader("Content-Type", contentType || "application/json");
+
+    if (isStream && upstreamRes.body) {
+        res.flushHeaders();
+        const reader = upstreamRes.body.getReader();
+        try {
+            while (!abortController.signal.aborted) {
+                const { done, value } = await reader.read();
+                if (done) break;
+                res.write(value);
+            }
+        } finally {
+            await reader.cancel().catch(() => {});
+            res.off("close", onClientClose);
+        }
+        if (!res.writableEnded) {
+            res.end();
+        }
+        return;
+    }
+
+    res.off("close", onClientClose);
+    const text = await upstreamRes.text();
+    res.send(text);
+}

server/routers/aiProvider/createAiModel.ts

@@ -0,0 +1,141 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiModels, aiProviders, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { and, eq } from "drizzle-orm";
+import type { CreateOrEditAiModelResponse } from "@server/routers/aiProvider/types";
+import { modelListTypeSchema } from "@server/lib/aiInferenceResource";
+
+const paramsSchema = z.strictObject({
+    providerId: z.coerce.number().int().positive()
+});
+
+const bodySchema = z.strictObject({
+    modelKey: z.string().nonempty(),
+    name: z.string().nonempty(),
+    enabled: z.boolean().optional(),
+    listType: modelListTypeSchema.optional().default("allow")
+});
+
+registry.registerPath({
+    method: "put",
+    path: "/ai-provider/{providerId}/model",
+    description: "Create an AI model under a provider.",
+    tags: [OpenAPITags.AiModel],
+    request: {
+        params: paramsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: bodySchema
+                }
+            }
+        }
+    },
+    responses: {
+        201: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function createAiModel(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const parsedBody = bodySchema.safeParse(req.body);
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { providerId } = parsedParams.data;
+        const { modelKey, name, enabled, listType } = parsedBody.data;
+
+        const [provider] =
+            req.aiProvider && req.aiProvider.providerId === providerId
+                ? [req.aiProvider]
+                : await db
+                      .select()
+                      .from(aiProviders)
+                      .where(eq(aiProviders.providerId, providerId))
+                      .limit(1);
+
+        if (!provider) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI provider with ID ${providerId} not found`
+                )
+            );
+        }
+
+        const [existing] = await db
+            .select({ modelId: aiModels.modelId })
+            .from(aiModels)
+            .where(
+                and(
+                    eq(aiModels.providerId, providerId),
+                    eq(aiModels.modelKey, modelKey)
+                )
+            )
+            .limit(1);
+
+        if (existing) {
+            return next(
+                createHttpError(
+                    HttpCode.CONFLICT,
+                    `Model with key ${modelKey} already exists for this provider`
+                )
+            );
+        }
+
+        const now = Date.now();
+        const [model] = await db
+            .insert(aiModels)
+            .values({
+                providerId,
+                modelKey,
+                name,
+                listType,
+                enabled: enabled ?? true,
+                createdAt: now,
+                updatedAt: now
+            })
+            .returning();
+
+        return response<CreateOrEditAiModelResponse>(res, {
+            data: { model },
+            success: true,
+            error: false,
+            message: "AI model created successfully",
+            status: HttpCode.CREATED
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiProvider/createAiProvider.ts

@@ -0,0 +1,171 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiProviders, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { encrypt } from "@server/lib/crypto";
+import config from "@server/lib/config";
+import {
+    resolveAiProviderCreateFields,
+    resolveCapabilitiesForCreate,
+    serializeAiProviderHeaders
+} from "@server/lib/aiProviderDefaults";
+import type { CreateOrEditAiProviderResponse } from "@server/routers/aiProvider/types";
+import { toPublicAiProvider } from "@server/routers/aiProvider/types";
+import {
+    aiAuthTypeSchema,
+    aiCapabilitiesSchema,
+    aiProviderHeadersSchema,
+    aiProviderTypeSchema,
+    aiRoutingModeSchema,
+    refineProviderUpstreamFields
+} from "@server/routers/aiProvider/validation";
+import { serializeCapabilities } from "@server/lib/aiCapabilities";
+
+const paramsSchema = z.strictObject({
+    orgId: z.string().nonempty()
+});
+
+const bodySchema = z
+    .strictObject({
+        name: z.string().nonempty(),
+        type: aiProviderTypeSchema,
+        upstreamUrl: z.url().optional().nullable(),
+        apiKey: z.string().optional(),
+        authType: aiAuthTypeSchema.optional(),
+        routingMode: aiRoutingModeSchema.optional(),
+        capabilities: aiCapabilitiesSchema.optional(),
+        headers: aiProviderHeadersSchema,
+        skipTlsVerification: z.boolean().optional(),
+        enabled: z.boolean().optional()
+    })
+    .superRefine((data, ctx) => {
+        refineProviderUpstreamFields(data, ctx);
+    });
+
+registry.registerPath({
+    method: "put",
+    path: "/org/{orgId}/ai-provider",
+    description: "Create an AI provider for an organization.",
+    tags: [OpenAPITags.AiProvider],
+    request: {
+        params: paramsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: bodySchema
+                }
+            }
+        }
+    },
+    responses: {
+        201: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function createAiProvider(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const parsedBody = bodySchema.safeParse(req.body);
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { orgId } = parsedParams.data;
+        const {
+            name,
+            type,
+            upstreamUrl,
+            apiKey,
+            authType,
+            routingMode,
+            capabilities,
+            headers,
+            skipTlsVerification,
+            enabled
+        } = parsedBody.data;
+
+        const key = config.getRawConfig().server.secret!;
+        const encryptedApiKey = apiKey ? encrypt(apiKey, key) : null;
+        const apiKeyLastChars = apiKey ? apiKey.slice(-4) : null;
+        const now = Date.now();
+        const resolved = resolveAiProviderCreateFields({
+            type,
+            upstreamUrl,
+            authType,
+            routingMode
+        });
+        const resolvedCapabilities = resolveCapabilitiesForCreate({
+            type,
+            capabilities
+        });
+
+        if (resolvedCapabilities.length === 0) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    "At least one capability is required"
+                )
+            );
+        }
+
+        const [provider] = await db
+            .insert(aiProviders)
+            .values({
+                orgId,
+                name,
+                type,
+                upstreamUrl: resolved.upstreamUrl,
+                apiKey: encryptedApiKey,
+                apiKeyLastChars,
+                authType: resolved.authType,
+                routingMode: resolved.routingMode,
+                capabilities: serializeCapabilities(resolvedCapabilities),
+                headers: serializeAiProviderHeaders(headers, key),
+                skipTlsVerification: skipTlsVerification ?? false,
+                enabled: enabled ?? true,
+                createdAt: now,
+                updatedAt: now
+            })
+            .returning();
+
+        return response<CreateOrEditAiProviderResponse>(res, {
+            data: {
+                provider: toPublicAiProvider(provider, { includeApiKey: true })
+            },
+            success: true,
+            error: false,
+            message: "AI provider created successfully",
+            status: HttpCode.CREATED
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiProvider/deleteAiModel.ts

@@ -0,0 +1,79 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiModels, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { eq } from "drizzle-orm";
+
+const paramsSchema = z.strictObject({
+    modelId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "delete",
+    path: "/ai-model/{modelId}",
+    description: "Delete an AI model.",
+    tags: [OpenAPITags.AiModel],
+    request: {
+        params: paramsSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function deleteAiModel(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { modelId } = parsedParams.data;
+
+        const [existing] = await db
+            .select({ modelId: aiModels.modelId })
+            .from(aiModels)
+            .where(eq(aiModels.modelId, modelId))
+            .limit(1);
+
+        if (!existing) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI model with ID ${modelId} not found`
+                )
+            );
+        }
+
+        await db.delete(aiModels).where(eq(aiModels.modelId, modelId));
+
+        return response(res, {
+            data: null,
+            success: true,
+            error: false,
+            message: "AI model deleted successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiProvider/deleteAiProvider.ts

@@ -0,0 +1,81 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiProviders, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { eq } from "drizzle-orm";
+
+const paramsSchema = z.strictObject({
+    providerId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "delete",
+    path: "/ai-provider/{providerId}",
+    description: "Delete an AI provider.",
+    tags: [OpenAPITags.AiProvider],
+    request: {
+        params: paramsSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function deleteAiProvider(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { providerId } = parsedParams.data;
+
+        const [existing] = await db
+            .select({ providerId: aiProviders.providerId })
+            .from(aiProviders)
+            .where(eq(aiProviders.providerId, providerId))
+            .limit(1);
+
+        if (!existing) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI provider with ID ${providerId} not found`
+                )
+            );
+        }
+
+        await db
+            .delete(aiProviders)
+            .where(eq(aiProviders.providerId, providerId));
+
+        return response(res, {
+            data: null,
+            success: true,
+            error: false,
+            message: "AI provider deleted successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiProvider/getAiModel.ts

@@ -0,0 +1,81 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiModels, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { eq } from "drizzle-orm";
+import type { GetAiModelResponse } from "@server/routers/aiProvider/types";
+
+const paramsSchema = z.strictObject({
+    modelId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "get",
+    path: "/ai-model/{modelId}",
+    description: "Get an AI model by ID.",
+    tags: [OpenAPITags.AiModel],
+    request: {
+        params: paramsSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function getAiModel(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { modelId } = parsedParams.data;
+
+        const [model] =
+            req.aiModel && req.aiModel.modelId === modelId
+                ? [req.aiModel]
+                : await db
+                      .select()
+                      .from(aiModels)
+                      .where(eq(aiModels.modelId, modelId))
+                      .limit(1);
+
+        if (!model) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI model with ID ${modelId} not found`
+                )
+            );
+        }
+
+        return response<GetAiModelResponse>(res, {
+            data: { model },
+            success: true,
+            error: false,
+            message: "AI model retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiProvider/getAiProvider.ts

@@ -0,0 +1,84 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiProviders, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { eq } from "drizzle-orm";
+import type { GetAiProviderResponse } from "@server/routers/aiProvider/types";
+import { toPublicAiProvider } from "@server/routers/aiProvider/types";
+
+const paramsSchema = z.strictObject({
+    providerId: z.coerce.number().int().positive()
+});
+
+registry.registerPath({
+    method: "get",
+    path: "/ai-provider/{providerId}",
+    description: "Get an AI provider by ID.",
+    tags: [OpenAPITags.AiProvider],
+    request: {
+        params: paramsSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function getAiProvider(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { providerId } = parsedParams.data;
+
+        const [provider] =
+            req.aiProvider && req.aiProvider.providerId === providerId
+                ? [req.aiProvider]
+                : await db
+                      .select()
+                      .from(aiProviders)
+                      .where(eq(aiProviders.providerId, providerId))
+                      .limit(1);
+
+        if (!provider) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI provider with ID ${providerId} not found`
+                )
+            );
+        }
+
+        return response<GetAiProviderResponse>(res, {
+            data: {
+                provider: toPublicAiProvider(provider, { includeApiKey: true })
+            },
+            success: true,
+            error: false,
+            message: "AI provider retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiProvider/index.ts

@@ -0,0 +1,12 @@
+export * from "./createAiProvider";
+export * from "./listAiProviders";
+export * from "./getAiProvider";
+export * from "./updateAiProvider";
+export * from "./deleteAiProvider";
+export * from "./createAiModel";
+export * from "./listAiModels";
+export * from "./listCatalogModels";
+export * from "./getAiModel";
+export * from "./updateAiModel";
+export * from "./deleteAiModel";
+export * from "./types";

server/routers/aiProvider/listAiModels.ts

@@ -0,0 +1,160 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiModels, aiProviders, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { and, asc, eq, like, sql } from "drizzle-orm";
+import type { ListAiModelsResponse } from "@server/routers/aiProvider/types";
+
+const paramsSchema = z.strictObject({
+    providerId: z.coerce.number().int().positive()
+});
+
+const listSchema = z.object({
+    pageSize: z.coerce
+        .number<string>()
+        .int()
+        .positive()
+        .optional()
+        .catch(20)
+        .default(20)
+        .openapi({
+            type: "integer",
+            default: 20,
+            description: "Number of items per page"
+        }),
+    page: z.coerce
+        .number<string>()
+        .int()
+        .min(0)
+        .optional()
+        .catch(1)
+        .default(1)
+        .openapi({
+            type: "integer",
+            default: 1,
+            description: "Page number to retrieve"
+        }),
+    query: z.string().optional()
+});
+
+registry.registerPath({
+    method: "get",
+    path: "/ai-provider/{providerId}/models",
+    description: "List AI models for a provider.",
+    tags: [OpenAPITags.AiModel],
+    request: {
+        params: paramsSchema,
+        query: listSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function listAiModels(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedQuery = listSchema.safeParse(req.query);
+        if (!parsedQuery.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedQuery.error).toString()
+                )
+            );
+        }
+
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { providerId } = parsedParams.data;
+
+        const [provider] =
+            req.aiProvider && req.aiProvider.providerId === providerId
+                ? [req.aiProvider]
+                : await db
+                      .select({ providerId: aiProviders.providerId })
+                      .from(aiProviders)
+                      .where(eq(aiProviders.providerId, providerId))
+                      .limit(1);
+
+        if (!provider) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI provider with ID ${providerId} not found`
+                )
+            );
+        }
+
+        const { pageSize, page, query } = parsedQuery.data;
+        const conditions = [eq(aiModels.providerId, providerId)];
+
+        if (query) {
+            conditions.push(
+                like(
+                    sql`LOWER(${aiModels.name})`,
+                    "%" + query.toLowerCase() + "%"
+                )
+            );
+        }
+
+        const baseQuery = db
+            .select()
+            .from(aiModels)
+            .where(and(...conditions));
+
+        const countQuery = db.$count(
+            db
+                .select()
+                .from(aiModels)
+                .where(and(...conditions))
+                .as("filtered_ai_models")
+        );
+
+        const [totalCount, rows] = await Promise.all([
+            countQuery,
+            baseQuery
+                .limit(pageSize)
+                .offset(pageSize * (page - 1))
+                .orderBy(asc(aiModels.name))
+        ]);
+
+        return response<ListAiModelsResponse>(res, {
+            data: {
+                models: rows,
+                pagination: {
+                    total: totalCount,
+                    pageSize,
+                    page
+                }
+            },
+            success: true,
+            error: false,
+            message: "AI models retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiProvider/listAiProviders.ts

@@ -0,0 +1,152 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiProviders, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { and, asc, eq, like, sql } from "drizzle-orm";
+import type { ListAiProvidersResponse } from "@server/routers/aiProvider/types";
+import { toPublicAiProvider } from "@server/routers/aiProvider/types";
+
+const paramsSchema = z.strictObject({
+    orgId: z.string().nonempty()
+});
+
+const listSchema = z.object({
+    pageSize: z.coerce
+        .number<string>()
+        .int()
+        .positive()
+        .optional()
+        .catch(20)
+        .default(20)
+        .openapi({
+            type: "integer",
+            default: 20,
+            description: "Number of items per page"
+        }),
+    page: z.coerce
+        .number<string>()
+        .int()
+        .min(0)
+        .optional()
+        .catch(1)
+        .default(1)
+        .openapi({
+            type: "integer",
+            default: 1,
+            description: "Page number to retrieve"
+        }),
+    query: z.string().optional()
+});
+
+registry.registerPath({
+    method: "get",
+    path: "/org/{orgId}/ai-providers",
+    description: "List AI providers for an organization.",
+    tags: [OpenAPITags.AiProvider],
+    request: {
+        params: paramsSchema,
+        query: listSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function listAiProviders(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedQuery = listSchema.safeParse(req.query);
+        if (!parsedQuery.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedQuery.error).toString()
+                )
+            );
+        }
+
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { orgId } = parsedParams.data;
+
+        if (req.user && orgId && orgId !== req.userOrgId) {
+            return next(
+                createHttpError(
+                    HttpCode.FORBIDDEN,
+                    "User does not have access to this organization"
+                )
+            );
+        }
+
+        const { pageSize, page, query } = parsedQuery.data;
+        const conditions = [eq(aiProviders.orgId, orgId)];
+
+        if (query) {
+            conditions.push(
+                like(
+                    sql`LOWER(${aiProviders.name})`,
+                    "%" + query.toLowerCase() + "%"
+                )
+            );
+        }
+
+        const baseQuery = db
+            .select()
+            .from(aiProviders)
+            .where(and(...conditions));
+
+        const countQuery = db.$count(
+            db
+                .select()
+                .from(aiProviders)
+                .where(and(...conditions))
+                .as("filtered_ai_providers")
+        );
+
+        const [totalCount, rows] = await Promise.all([
+            countQuery,
+            baseQuery
+                .limit(pageSize)
+                .offset(pageSize * (page - 1))
+                .orderBy(asc(aiProviders.name))
+        ]);
+
+        return response<ListAiProvidersResponse>(res, {
+            data: {
+                providers: rows.map((row) => toPublicAiProvider(row)),
+                pagination: {
+                    total: totalCount,
+                    pageSize,
+                    page
+                }
+            },
+            success: true,
+            error: false,
+            message: "AI providers retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiProvider/listCatalogModels.ts

@@ -0,0 +1,131 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiProviders, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { eq } from "drizzle-orm";
+import {
+    aiModelCatalog,
+    getCatalogProviderForType
+} from "@server/lib/aiModelCatalog";
+import type { AiProviderType } from "@server/lib/aiProviderDefaults";
+import type { ListCatalogModelsResponse } from "@server/routers/aiProvider/types";
+
+const paramsSchema = z.strictObject({
+    providerId: z.coerce.number().int().positive()
+});
+
+const listSchema = z.object({
+    query: z.string().optional()
+});
+
+registry.registerPath({
+    method: "get",
+    path: "/ai-provider/{providerId}/catalog-models",
+    description:
+        "List known catalog models for an AI provider's type. Used for model key suggestions.",
+    tags: [OpenAPITags.AiModel],
+    request: {
+        params: paramsSchema,
+        query: listSchema
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function listCatalogModels(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedQuery = listSchema.safeParse(req.query);
+        if (!parsedQuery.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedQuery.error).toString()
+                )
+            );
+        }
+
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const { providerId } = parsedParams.data;
+
+        const [provider] =
+            req.aiProvider && req.aiProvider.providerId === providerId
+                ? [req.aiProvider]
+                : await db
+                      .select()
+                      .from(aiProviders)
+                      .where(eq(aiProviders.providerId, providerId))
+                      .limit(1);
+
+        if (!provider) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI provider with ID ${providerId} not found`
+                )
+            );
+        }
+
+        const catalogProvider = getCatalogProviderForType(
+            provider.type as AiProviderType
+        );
+
+        let models = catalogProvider
+            ? aiModelCatalog.list(catalogProvider).map((entry) => ({
+                  model: entry.model
+              }))
+            : [];
+
+        const { query } = parsedQuery.data;
+        if (query) {
+            const q = query.toLowerCase();
+            models = models.filter((m) => m.model.toLowerCase().includes(q));
+        }
+
+        // Deduplicate model keys (catalog may have duplicates after provider
+        // normalization, e.g. bedrock + bedrock_converse).
+        const seen = new Set<string>();
+        models = models.filter((m) => {
+            if (seen.has(m.model)) {
+                return false;
+            }
+            seen.add(m.model);
+            return true;
+        });
+
+        models.sort((a, b) => a.model.localeCompare(b.model));
+
+        return response<ListCatalogModelsResponse>(res, {
+            data: { models },
+            success: true,
+            error: false,
+            message: "Catalog models retrieved successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiProvider/types.ts

@@ -0,0 +1,90 @@
+import type { AiModel, AiProvider } from "@server/db";
+import type { PaginatedResponse } from "@server/types/Pagination";
+import {
+    parseAiProviderHeaders,
+    type AiProviderAuthType,
+    type AiProviderHeader
+} from "@server/lib/aiProviderDefaults";
+import {
+    parseCapabilities,
+    type AiCapability
+} from "@server/lib/aiCapabilities";
+import { decrypt } from "@server/lib/crypto";
+import config from "@server/lib/config";
+
+export type AiProviderPublic = Omit<
+    AiProvider,
+    "apiKey" | "capabilities" | "headers"
+> & {
+    apiKey?: string | null;
+    capabilities: AiCapability[];
+    headers: AiProviderHeader[] | null;
+    effectiveUpstreamUrl: string | null;
+    effectiveAuthType: AiProviderAuthType;
+};
+
+export type ListAiProvidersResponse = PaginatedResponse<{
+    providers: AiProviderPublic[];
+}>;
+
+export type GetAiProviderResponse = {
+    provider: AiProviderPublic;
+};
+
+export type CreateOrEditAiProviderResponse = {
+    provider: AiProviderPublic;
+};
+
+export type ListAiModelsResponse = PaginatedResponse<{
+    models: AiModel[];
+}>;
+
+export type ListCatalogModelsResponse = {
+    models: { model: string }[];
+};
+
+export type GetAiModelResponse = {
+    model: AiModel;
+};
+
+export type CreateOrEditAiModelResponse = {
+    model: AiModel;
+};
+
+export function toPublicAiProvider(
+    provider: AiProvider,
+    options?: { includeApiKey?: boolean }
+): AiProviderPublic {
+    const {
+        apiKey: encryptedApiKey,
+        capabilities: rawCapabilities,
+        headers: rawHeaders,
+        ...rest
+    } = provider;
+
+    let apiKey: string | null | undefined;
+    if (options?.includeApiKey) {
+        if (encryptedApiKey) {
+            apiKey = decrypt(
+                encryptedApiKey,
+                config.getRawConfig().server.secret!
+            );
+        } else {
+            apiKey = null;
+        }
+    }
+
+    const parsedHeaders = parseAiProviderHeaders(
+        rawHeaders,
+        config.getRawConfig().server.secret!
+    );
+
+    return {
+        ...rest,
+        ...(options?.includeApiKey ? { apiKey } : {}),
+        capabilities: parseCapabilities(rawCapabilities),
+        headers: parsedHeaders.length > 0 ? parsedHeaders : null,
+        effectiveUpstreamUrl: provider.upstreamUrl,
+        effectiveAuthType: provider.authType as AiProviderAuthType
+    };
+}

server/routers/aiProvider/updateAiModel.ts

@@ -0,0 +1,156 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiModels, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { and, eq, ne } from "drizzle-orm";
+import type { CreateOrEditAiModelResponse } from "@server/routers/aiProvider/types";
+import { modelListTypeSchema } from "@server/lib/aiInferenceResource";
+
+const paramsSchema = z.strictObject({
+    modelId: z.coerce.number().int().positive()
+});
+
+const bodySchema = z.strictObject({
+    modelKey: z.string().nonempty().optional(),
+    name: z.string().nonempty().optional(),
+    enabled: z.boolean().optional(),
+    listType: modelListTypeSchema.optional()
+});
+
+registry.registerPath({
+    method: "post",
+    path: "/ai-model/{modelId}",
+    description: "Update an AI model.",
+    tags: [OpenAPITags.AiModel],
+    request: {
+        params: paramsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: bodySchema
+                }
+            }
+        }
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function updateAiModel(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const parsedBody = bodySchema.safeParse(req.body);
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { modelId } = parsedParams.data;
+        const body = parsedBody.data;
+
+        const [existing] =
+            req.aiModel && req.aiModel.modelId === modelId
+                ? [req.aiModel]
+                : await db
+                      .select()
+                      .from(aiModels)
+                      .where(eq(aiModels.modelId, modelId))
+                      .limit(1);
+
+        if (!existing) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI model with ID ${modelId} not found`
+                )
+            );
+        }
+
+        if (
+            body.modelKey !== undefined &&
+            body.modelKey !== existing.modelKey
+        ) {
+            const [conflict] = await db
+                .select({ modelId: aiModels.modelId })
+                .from(aiModels)
+                .where(
+                    and(
+                        eq(aiModels.providerId, existing.providerId),
+                        eq(aiModels.modelKey, body.modelKey),
+                        ne(aiModels.modelId, modelId)
+                    )
+                )
+                .limit(1);
+
+            if (conflict) {
+                return next(
+                    createHttpError(
+                        HttpCode.CONFLICT,
+                        `Model with key ${body.modelKey} already exists for this provider`
+                    )
+                );
+            }
+        }
+
+        const updateData: Partial<typeof aiModels.$inferInsert> = {
+            updatedAt: Date.now()
+        };
+
+        if (body.modelKey !== undefined) {
+            updateData.modelKey = body.modelKey;
+        }
+        if (body.name !== undefined) {
+            updateData.name = body.name;
+        }
+        if (body.enabled !== undefined) {
+            updateData.enabled = body.enabled;
+        }
+        if (body.listType !== undefined) {
+            updateData.listType = body.listType;
+        }
+
+        const [model] = await db
+            .update(aiModels)
+            .set(updateData)
+            .where(eq(aiModels.modelId, modelId))
+            .returning();
+
+        return response<CreateOrEditAiModelResponse>(res, {
+            data: { model },
+            success: true,
+            error: false,
+            message: "AI model updated successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiProvider/updateAiProvider.ts

@@ -0,0 +1,223 @@
+import { Request, Response, NextFunction } from "express";
+import { z } from "zod";
+import { aiProviders, db } from "@server/db";
+import response from "@server/lib/response";
+import HttpCode from "@server/types/HttpCode";
+import createHttpError from "http-errors";
+import logger from "@server/logger";
+import { fromError } from "zod-validation-error";
+import { OpenAPITags, registry } from "@server/openApi";
+import { eq } from "drizzle-orm";
+import { encrypt } from "@server/lib/crypto";
+import config from "@server/lib/config";
+import type { CreateOrEditAiProviderResponse } from "@server/routers/aiProvider/types";
+import { toPublicAiProvider } from "@server/routers/aiProvider/types";
+import {
+    aiAuthTypeSchema,
+    aiCapabilitiesSchema,
+    aiProviderHeadersSchema,
+    aiProviderTypeSchema,
+    aiRoutingModeSchema,
+    refineProviderUpstreamFields
+} from "@server/routers/aiProvider/validation";
+import {
+    serializeAiProviderHeaders,
+    type AiProviderAuthType,
+    type AiProviderRoutingMode,
+    type AiProviderType
+} from "@server/lib/aiProviderDefaults";
+import {
+    parseCapabilities,
+    serializeCapabilities
+} from "@server/lib/aiCapabilities";
+
+const paramsSchema = z.strictObject({
+    providerId: z.coerce.number().int().positive()
+});
+
+const bodySchema = z.strictObject({
+    name: z.string().nonempty().optional(),
+    upstreamUrl: z.url().optional().nullable(),
+    apiKey: z.string().optional(),
+    authType: aiAuthTypeSchema.optional(),
+    routingMode: aiRoutingModeSchema.optional(),
+    capabilities: aiCapabilitiesSchema.optional(),
+    headers: aiProviderHeadersSchema,
+    skipTlsVerification: z.boolean().optional(),
+    enabled: z.boolean().optional()
+});
+
+registry.registerPath({
+    method: "post",
+    path: "/ai-provider/{providerId}",
+    description: "Update an AI provider.",
+    tags: [OpenAPITags.AiProvider],
+    request: {
+        params: paramsSchema,
+        body: {
+            content: {
+                "application/json": {
+                    schema: bodySchema
+                }
+            }
+        }
+    },
+    responses: {
+        200: {
+            description: "Successful response"
+        }
+    }
+});
+
+export async function updateAiProvider(
+    req: Request,
+    res: Response,
+    next: NextFunction
+): Promise<any> {
+    try {
+        const parsedParams = paramsSchema.safeParse(req.params);
+        if (!parsedParams.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedParams.error).toString()
+                )
+            );
+        }
+
+        const parsedBody = bodySchema.safeParse(req.body);
+        if (!parsedBody.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(parsedBody.error).toString()
+                )
+            );
+        }
+
+        const { providerId } = parsedParams.data;
+        const body = parsedBody.data;
+
+        const [existing] =
+            req.aiProvider && req.aiProvider.providerId === providerId
+                ? [req.aiProvider]
+                : await db
+                      .select()
+                      .from(aiProviders)
+                      .where(eq(aiProviders.providerId, providerId))
+                      .limit(1);
+
+        if (!existing) {
+            return next(
+                createHttpError(
+                    HttpCode.NOT_FOUND,
+                    `AI provider with ID ${providerId} not found`
+                )
+            );
+        }
+
+        const providerType = existing.type as AiProviderType;
+        const nextRoutingMode: AiProviderRoutingMode =
+            providerType === "custom"
+                ? ((body.routingMode ??
+                      existing.routingMode) as AiProviderRoutingMode)
+                : "url";
+        const nextUpstreamUrl =
+            body.upstreamUrl !== undefined
+                ? body.upstreamUrl
+                : existing.upstreamUrl;
+        const nextAuthType: AiProviderAuthType =
+            body.authType !== undefined
+                ? body.authType
+                : (existing.authType as AiProviderAuthType);
+
+        const nextCapabilities =
+            body.capabilities !== undefined
+                ? body.capabilities
+                : parseCapabilities(existing.capabilities);
+
+        const validation = z
+            .object({
+                type: aiProviderTypeSchema,
+                upstreamUrl: z.string().nullable().optional(),
+                authType: aiAuthTypeSchema,
+                routingMode: aiRoutingModeSchema.optional(),
+                capabilities: aiCapabilitiesSchema.optional()
+            })
+            .superRefine((data, ctx) => refineProviderUpstreamFields(data, ctx))
+            .safeParse({
+                type: providerType,
+                upstreamUrl: nextUpstreamUrl,
+                authType: nextAuthType,
+                routingMode: nextRoutingMode,
+                capabilities: nextCapabilities
+            });
+
+        if (!validation.success) {
+            return next(
+                createHttpError(
+                    HttpCode.BAD_REQUEST,
+                    fromError(validation.error).toString()
+                )
+            );
+        }
+
+        const updateData: Partial<typeof aiProviders.$inferInsert> = {
+            updatedAt: Date.now(),
+            routingMode: nextRoutingMode
+        };
+
+        if (body.name !== undefined) {
+            updateData.name = body.name;
+        }
+        if (body.skipTlsVerification !== undefined) {
+            updateData.skipTlsVerification = body.skipTlsVerification;
+        }
+        if (body.enabled !== undefined) {
+            updateData.enabled = body.enabled;
+        }
+        if (nextRoutingMode === "target") {
+            updateData.upstreamUrl = null;
+        } else if (body.upstreamUrl !== undefined) {
+            updateData.upstreamUrl = body.upstreamUrl;
+        }
+        if (body.authType !== undefined) {
+            updateData.authType = body.authType;
+        }
+        if (body.capabilities !== undefined) {
+            updateData.capabilities = serializeCapabilities(body.capabilities);
+        }
+
+        if (body.apiKey !== undefined) {
+            const key = config.getRawConfig().server.secret!;
+            updateData.apiKey = encrypt(body.apiKey, key);
+            updateData.apiKeyLastChars = body.apiKey.slice(-4);
+        }
+
+        if (body.headers !== undefined) {
+            const key = config.getRawConfig().server.secret!;
+            updateData.headers = serializeAiProviderHeaders(body.headers, key);
+        }
+
+        const [provider] = await db
+            .update(aiProviders)
+            .set(updateData)
+            .where(eq(aiProviders.providerId, providerId))
+            .returning();
+
+        return response<CreateOrEditAiProviderResponse>(res, {
+            data: {
+                provider: toPublicAiProvider(provider, { includeApiKey: true })
+            },
+            success: true,
+            error: false,
+            message: "AI provider updated successfully",
+            status: HttpCode.OK
+        });
+    } catch (error) {
+        logger.error(error);
+        return next(
+            createHttpError(HttpCode.INTERNAL_SERVER_ERROR, "An error occurred")
+        );
+    }
+}

server/routers/aiProvider/validation.ts

@@ -0,0 +1,126 @@
+import { z } from "zod";
+import {
+    AI_PROVIDER_AUTH_TYPES,
+    providerRequiresUpstreamUrl,
+    type AiProviderAuthType,
+    type AiProviderRoutingMode,
+    type AiProviderType
+} from "@server/lib/aiProviderDefaults";
+import { AI_CAPABILITIES } from "@server/lib/aiCapabilities";
+
+export const aiProviderTypeSchema = z.enum([
+    "openai",
+    "anthropic",
+    "googleGemini",
+    "vertexAi",
+    "bedrock",
+    "microsoftFoundry",
+    "openRouter",
+    "vercelAiGateway",
+    "custom"
+]);
+
+export const aiAuthTypeSchema = z.enum(AI_PROVIDER_AUTH_TYPES);
+
+export const aiRoutingModeSchema = z.enum(["url", "target"]);
+
+export const aiCapabilitySchema = z.enum(AI_CAPABILITIES);
+
+export const aiCapabilitiesSchema = z.array(aiCapabilitySchema);
+
+const validHeaderName = /^[a-zA-Z0-9!#$%&'*+\-.^_`|~]+$/;
+const validHeaderValue = /^[\t\x20-\x7E]*$/;
+const templatePattern = /\{\{[^}]+\}\}/;
+
+export const aiProviderHeadersSchema = z
+    .array(z.strictObject({ name: z.string(), value: z.string() }))
+    .nullable()
+    .optional()
+    .superRefine((headers, ctx) => {
+        if (!headers) {
+            return;
+        }
+        for (const [index, header] of headers.entries()) {
+            if (!validHeaderName.test(header.name)) {
+                ctx.addIssue({
+                    code: "custom",
+                    message:
+                        "Header names may only contain valid HTTP token characters (letters, digits, and !#$%&'*+-.^_`|~).",
+                    path: [index, "name"]
+                });
+            }
+            if (!validHeaderValue.test(header.value)) {
+                ctx.addIssue({
+                    code: "custom",
+                    message:
+                        "Header values may only contain printable ASCII characters and horizontal whitespace.",
+                    path: [index, "value"]
+                });
+            }
+            if (
+                templatePattern.test(header.name) ||
+                templatePattern.test(header.value)
+            ) {
+                ctx.addIssue({
+                    code: "custom",
+                    message:
+                        "Header names and values must not contain template expressions such as {{value}}.",
+                    path: [index]
+                });
+            }
+        }
+    });
+
+export function refineProviderUpstreamFields(
+    data: {
+        type: AiProviderType;
+        upstreamUrl?: string | null;
+        authType?: AiProviderAuthType | null;
+        routingMode?: AiProviderRoutingMode | null;
+        capabilities?: z.infer<typeof aiCapabilitiesSchema> | null;
+    },
+    ctx: z.RefinementCtx
+) {
+    const routingMode = data.routingMode ?? "url";
+
+    if (data.type !== "custom" && routingMode === "target") {
+        ctx.addIssue({
+            code: "custom",
+            message: "routingMode target is only allowed for custom providers",
+            path: ["routingMode"]
+        });
+    }
+
+    if (
+        providerRequiresUpstreamUrl(data.type, routingMode) &&
+        !data.upstreamUrl
+    ) {
+        ctx.addIssue({
+            code: "custom",
+            message: `upstreamUrl is required for ${data.type} providers`,
+            path: ["upstreamUrl"]
+        });
+    }
+
+    if (data.type === "custom") {
+        const caps = data.capabilities;
+        if (!caps || caps.length === 0) {
+            ctx.addIssue({
+                code: "custom",
+                message:
+                    "At least one capability is required for custom providers",
+                path: ["capabilities"]
+            });
+        }
+    } else if (
+        data.capabilities !== undefined &&
+        data.capabilities !== null &&
+        data.capabilities.length === 0
+    ) {
+        ctx.addIssue({
+            code: "custom",
+            message: "At least one capability is required",
+            path: ["capabilities"]
+        });
+    }
+}

server/routers/badger/verifySession.ts

@@ -222,7 +222,9 @@ export async function verifyResourceSession(
         }
 
         const { blockAccess, mode } = resource;
-        const dontStripSession = ["ssh", "rdp", "vnc"].includes(mode);
+        const dontStripSession = ["ssh", "rdp", "vnc", "inference"].includes(
+            mode
+        );
 
         if (blockAccess) {
             logger.debug("Resource blocked", host);

server/routers/client/createClient.ts

@@ -255,20 +255,6 @@ export async function createClient(
 
         let newClient: Client | null = null;
         await db.transaction(async (trx) => {
-            // TODO: more intelligent way to pick the exit node
-            const exitNodesList = await listExitNodes(orgId);
-            const randomExitNode =
-                exitNodesList[Math.floor(Math.random() * exitNodesList.length)];
-
-            if (!randomExitNode) {
-                return next(
-                    createHttpError(
-                        HttpCode.NOT_FOUND,
-                        `No exit nodes available. ${build == "saas" ? "Please contact support." : "You need to install gerbil to use the clients."}`
-                    )
-                );
-            }
-
             const [adminRole] = await trx
                 .select()
                 .from(roles)
@@ -287,7 +273,6 @@ export async function createClient(
                 .insert(clients)
                 .values({
                     niceId,
-                    exitNodeId: randomExitNode.exitNodeId,
                     orgId,
                     name,
                     subnet: updatedSubnet,

server/routers/client/createUserClient.ts

@@ -222,11 +222,6 @@ export async function createUserClient(
 
         let newClient: Client | null = null;
         await db.transaction(async (trx) => {
-            // TODO: more intelligent way to pick the exit node
-            const exitNodesList = await listExitNodes(orgId);
-            const randomExitNode =
-                exitNodesList[Math.floor(Math.random() * exitNodesList.length)];
-
             const [adminRole] = await trx
                 .select()
                 .from(roles)
@@ -244,7 +239,6 @@ export async function createUserClient(
             [newClient] = await trx
                 .insert(clients)
                 .values({
-                    exitNodeId: randomExitNode.exitNodeId,
                     orgId,
                     niceId,
                     name,

server/routers/external.ts

@@ -45,7 +45,10 @@ import {
     verifySiteResourceAccess,
     verifyOlmAccess,
     verifyLimits,
-    verifyResourcePolicyAccess
+    verifyResourcePolicyAccess,
+    verifyAiProviderAccess,
+    verifyAiModelAccess,
+    verifyAiBudgetAccess
 } from "@server/middlewares";
 import { ActionsEnum } from "@server/auth/actions";
 import rateLimit, { ipKeyGenerator } from "express-rate-limit";
@@ -55,6 +58,8 @@ import { createStore } from "#dynamic/lib/rateLimitStore";
 import { logActionAudit } from "#dynamic/middlewares";
 import { checkRoundTripMessage } from "./ws";
 import * as labels from "@server/routers/labels";
+import * as aiProvider from "@server/routers/aiProvider";
+import * as aiBudget from "@server/routers/aiBudget";
 
 // Root routes
 export const unauthenticated = Router();
@@ -404,6 +409,20 @@ authenticated.get(
     siteResource.listSiteResourceClients
 );
 
+authenticated.get(
+    "/site-resource/:siteResourceId/ai-models",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.listResourceAiModels),
+    siteResource.listSiteResourceAiModels
+);
+
+authenticated.get(
+    "/site-resource/:siteResourceId/ai-providers",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.listResourceAiModels),
+    siteResource.listSiteResourceAiProviders
+);
+
 authenticated.post(
     "/site-resource/:siteResourceId/roles",
     verifySiteResourceAccess,
@@ -414,6 +433,54 @@ authenticated.post(
     siteResource.setSiteResourceRoles
 );
 
+authenticated.post(
+    "/site-resource/:siteResourceId/ai-models",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.setSiteResourceAiModels
+);
+
+authenticated.post(
+    "/site-resource/:siteResourceId/ai-models/add",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.addAiModelToSiteResource
+);
+
+authenticated.post(
+    "/site-resource/:siteResourceId/ai-models/remove",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.removeAiModelFromSiteResource
+);
+
+authenticated.post(
+    "/site-resource/:siteResourceId/ai-providers",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.setSiteResourceAiProviders
+);
+
+authenticated.post(
+    "/site-resource/:siteResourceId/ai-providers/add",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.addAiProviderToSiteResource
+);
+
+authenticated.post(
+    "/site-resource/:siteResourceId/ai-providers/remove",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.removeAiProviderFromSiteResource
+);
+
 authenticated.post(
     "/site-resource/:siteResourceId/users",
     verifySiteResourceAccess,
@@ -648,6 +715,20 @@ authenticated.get(
     resource.listResourceUsers
 );
 
+authenticated.get(
+    "/resource/:resourceId/ai-models",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.listResourceAiModels),
+    resource.listResourceAiModels
+);
+
+authenticated.get(
+    "/resource/:resourceId/ai-providers",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.listResourceAiModels),
+    resource.listResourceAiProviders
+);
+
 authenticated.get(
     "/resource/:resourceId",
     verifyResourceAccess,
@@ -851,6 +932,54 @@ authenticated.post(
     resource.setResourceUsers
 );
 
+authenticated.post(
+    "/resource/:resourceId/ai-models",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.setResourceAiModels
+);
+
+authenticated.post(
+    "/resource/:resourceId/ai-models/add",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.addAiModelToResource
+);
+
+authenticated.post(
+    "/resource/:resourceId/ai-models/remove",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.removeAiModelFromResource
+);
+
+authenticated.post(
+    "/resource/:resourceId/ai-providers",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.setResourceAiProviders
+);
+
+authenticated.post(
+    "/resource/:resourceId/ai-providers/add",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.addAiProviderToResource
+);
+
+authenticated.post(
+    "/resource/:resourceId/ai-providers/remove",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.removeAiProviderFromResource
+);
+
 authenticated.put(
     "/resource-policy/:resourcePolicyId/access-control",
     verifyResourcePolicyAccess,
@@ -1366,6 +1495,179 @@ authenticated.get(
 
 authenticated.get("/ws/round-trip-message/:messageId", checkRoundTripMessage);
 
+authenticated.put(
+    "/org/:orgId/ai-provider",
+    verifyOrgAccess,
+    verifyUserHasAction(ActionsEnum.createAiProvider),
+    logActionAudit(ActionsEnum.createAiProvider),
+    aiProvider.createAiProvider
+);
+
+authenticated.get(
+    "/org/:orgId/ai-providers",
+    verifyOrgAccess,
+    verifyUserHasAction(ActionsEnum.listAiProviders),
+    aiProvider.listAiProviders
+);
+
+authenticated.get(
+    "/ai-provider/:providerId",
+    verifyAiProviderAccess,
+    verifyUserHasAction(ActionsEnum.getAiProvider),
+    aiProvider.getAiProvider
+);
+
+authenticated.put(
+    "/ai-provider/:providerId/target",
+    verifyAiProviderAccess,
+    verifySiteAccess,
+    verifyLimits,
+    verifyUserHasAction(ActionsEnum.createTarget),
+    logActionAudit(ActionsEnum.createTarget),
+    target.createTarget
+);
+
+authenticated.get(
+    "/ai-provider/:providerId/targets",
+    verifyAiProviderAccess,
+    verifyUserHasAction(ActionsEnum.listTargets),
+    target.listTargets
+);
+
+authenticated.post(
+    "/ai-provider/:providerId",
+    verifyAiProviderAccess,
+    verifyUserHasAction(ActionsEnum.updateAiProvider),
+    logActionAudit(ActionsEnum.updateAiProvider),
+    aiProvider.updateAiProvider
+);
+
+authenticated.delete(
+    "/ai-provider/:providerId",
+    verifyAiProviderAccess,
+    verifyUserHasAction(ActionsEnum.deleteAiProvider),
+    logActionAudit(ActionsEnum.deleteAiProvider),
+    aiProvider.deleteAiProvider
+);
+
+authenticated.put(
+    "/ai-provider/:providerId/model",
+    verifyAiProviderAccess,
+    verifyUserHasAction(ActionsEnum.createAiModel),
+    logActionAudit(ActionsEnum.createAiModel),
+    aiProvider.createAiModel
+);
+
+authenticated.get(
+    "/ai-provider/:providerId/models",
+    verifyAiProviderAccess,
+    verifyUserHasAction(ActionsEnum.listAiModels),
+    aiProvider.listAiModels
+);
+
+authenticated.get(
+    "/ai-provider/:providerId/catalog-models",
+    verifyAiProviderAccess,
+    verifyUserHasAction(ActionsEnum.listAiModels),
+    aiProvider.listCatalogModels
+);
+
+authenticated.get(
+    "/ai-model/:modelId",
+    verifyAiModelAccess,
+    verifyUserHasAction(ActionsEnum.getAiModel),
+    aiProvider.getAiModel
+);
+
+authenticated.post(
+    "/ai-model/:modelId",
+    verifyAiModelAccess,
+    verifyUserHasAction(ActionsEnum.updateAiModel),
+    logActionAudit(ActionsEnum.updateAiModel),
+    aiProvider.updateAiModel
+);
+
+authenticated.delete(
+    "/ai-model/:modelId",
+    verifyAiModelAccess,
+    verifyUserHasAction(ActionsEnum.deleteAiModel),
+    logActionAudit(ActionsEnum.deleteAiModel),
+    aiProvider.deleteAiModel
+);
+
+authenticated.put(
+    "/org/:orgId/ai-budget",
+    verifyOrgAccess,
+    verifyUserHasAction(ActionsEnum.createAiBudget),
+    logActionAudit(ActionsEnum.createAiBudget),
+    aiBudget.createAiBudget
+);
+
+authenticated.get(
+    "/org/:orgId/ai-budgets",
+    verifyOrgAccess,
+    verifyUserHasAction(ActionsEnum.listAiBudgets),
+    aiBudget.listAiBudgets
+);
+
+authenticated.get(
+    "/ai-budget/:budgetId",
+    verifyAiBudgetAccess,
+    verifyUserHasAction(ActionsEnum.getAiBudget),
+    aiBudget.getAiBudget
+);
+
+authenticated.post(
+    "/ai-budget/:budgetId",
+    verifyAiBudgetAccess,
+    verifyUserHasAction(ActionsEnum.updateAiBudget),
+    logActionAudit(ActionsEnum.updateAiBudget),
+    aiBudget.updateAiBudget
+);
+
+authenticated.delete(
+    "/ai-budget/:budgetId",
+    verifyAiBudgetAccess,
+    verifyUserHasAction(ActionsEnum.deleteAiBudget),
+    logActionAudit(ActionsEnum.deleteAiBudget),
+    aiBudget.deleteAiBudget
+);
+
+authenticated.get(
+    "/ai-provider/:providerId/ai-budgets",
+    verifyAiProviderAccess,
+    verifyUserHasAction(ActionsEnum.listAiBudgets),
+    aiBudget.listAiBudgetsForProvider
+);
+
+authenticated.get(
+    "/ai-model/:modelId/ai-budgets",
+    verifyAiModelAccess,
+    verifyUserHasAction(ActionsEnum.listAiBudgets),
+    aiBudget.listAiBudgetsForModel
+);
+
+authenticated.get(
+    "/resource/:resourceId/ai-budgets",
+    verifyResourceAccess,
+    verifyUserHasAction(ActionsEnum.listAiBudgets),
+    aiBudget.listAiBudgetsForResource
+);
+
+authenticated.get(
+    "/site-resource/:siteResourceId/ai-budgets",
+    verifySiteResourceAccess,
+    verifyUserHasAction(ActionsEnum.listAiBudgets),
+    aiBudget.listAiBudgetsForSiteResource
+);
+
+authenticated.get(
+    "/role/:roleId/ai-budgets",
+    verifyRoleAccess,
+    verifyUserHasAction(ActionsEnum.listAiBudgets),
+    aiBudget.listAiBudgetsForRole
+);
+
 authenticated.get(
     "/org/:orgId/labels",
     verifyOrgAccess,

server/routers/gerbil/getAllRelays.ts

@@ -100,7 +100,7 @@ export async function generateRelayMappings(exitNode: ExitNode) {
     // Filter to sites with the required fields up front so the rest of the
     // function can safely treat endpoint/subnet/listenPort as defined.
     const validSites = sitesRes.filter(
-        (s) => s.endpoint && s.subnet && s.listenPort
+        (s) => s.endpoint && s.exitNodeSubnet && s.listenPort
     );
 
     if (validSites.length === 0) {
@@ -136,7 +136,7 @@ export async function generateRelayMappings(exitNode: ExitNode) {
         if (
             peer.orgId == null ||
             !peer.endpoint ||
-            !peer.subnet ||
+            !peer.exitNodeSubnet ||
             !peer.listenPort
         ) {
             continue;
@@ -183,7 +183,7 @@ export async function generateRelayMappings(exitNode: ExitNode) {
     // Process each site using the pre-fetched data.
     for (const site of validSites) {
         const siteDestination: PeerDestination = {
-            destinationIP: site.subnet!.split("/")[0],
+            destinationIP: site.exitNodeSubnet!.split("/")[0],
             destinationPort: site.listenPort! || 1 // this satisfies gerbil for now but should be reevaluated
         };
 
@@ -207,7 +207,7 @@ export async function generateRelayMappings(exitNode: ExitNode) {
                         continue;
                     }
                     addDestination(site.endpoint!, {
-                        destinationIP: peer.subnet!.split("/")[0],
+                        destinationIP: peer.exitNodeSubnet!.split("/")[0],
                         destinationPort: peer.listenPort! || 1 // this satisfies gerbil for now but should be reevaluated
                     });
                 }

server/routers/gerbil/getConfig.ts

@@ -1,6 +1,6 @@
 import { Request, Response, NextFunction } from "express";
 import { z } from "zod";
-import { sites, exitNodes, ExitNode } from "@server/db";
+import { sites, exitNodes, ExitNode, clients } from "@server/db";
 import { db } from "@server/db";
 import { eq, isNotNull, and } from "drizzle-orm";
 import HttpCode from "@server/types/HttpCode";
@@ -89,11 +89,27 @@ export async function generateGerbilConfig(exitNode: ExitNode) {
             and(
                 eq(sites.exitNodeId, exitNode.exitNodeId),
                 isNotNull(sites.pubKey),
-                isNotNull(sites.subnet)
+                isNotNull(sites.exitNodeSubnet)
             )
         );
 
-    const peers = await Promise.all(
+    const clientsRes = await db
+        .select()
+        .from(clients)
+        .where(
+            and(
+                eq(clients.exitNodeId, exitNode.exitNodeId),
+                isNotNull(clients.pubKey),
+                isNotNull(clients.exitNodeSubnet)
+            )
+        );
+
+    let peers: {
+        publicKey: string | null;
+        allowedIps: string[];
+    }[] = [];
+
+    const sitePeers = await Promise.all(
         sitesRes.map(async (site) => {
             if (site.type === "wireguard") {
                 return {
@@ -103,7 +119,7 @@ export async function generateGerbilConfig(exitNode: ExitNode) {
             } else if (site.type === "newt") {
                 return {
                     publicKey: site.pubKey,
-                    allowedIps: [site.subnet!]
+                    allowedIps: [site.exitNodeSubnet!]
                 };
             }
             return {
@@ -113,6 +129,15 @@ export async function generateGerbilConfig(exitNode: ExitNode) {
         })
     );
 
+    const clientPeers = clientsRes.map((client) => {
+        return {
+            publicKey: client.pubKey,
+            allowedIps: [client.exitNodeSubnet!]
+        };
+    });
+
+    peers = [...sitePeers, ...clientPeers];
+
     const configResponse: GetConfigResponse = {
         listenPort: exitNode.listenPort || 51820,
         ipAddress: exitNode.address,

server/routers/gerbil/updateHolePunch.ts

@@ -188,7 +188,7 @@ export async function updateAndGenerateEndpointDestinations(
             .select({
                 siteId: sites.siteId,
                 newtId: newts.newtId,
-                subnet: sites.subnet,
+                subnet: sites.exitNodeSubnet,
                 listenPort: sites.listenPort,
                 publicKey: sites.publicKey,
                 endpoint: clientSitesAssociationsCache.endpoint,

server/routers/integration.ts

@@ -13,6 +13,7 @@ import * as apiKeys from "./apiKeys";
 import * as idp from "./idp";
 import * as logs from "./auditLogs";
 import * as siteResource from "./siteResource";
+import * as aiProvider from "./aiProvider";
 import {
     verifyApiKey,
     verifyApiKeyOrgAccess,
@@ -31,6 +32,8 @@ import {
     verifyLimits,
     verifyApiKeyDomainAccess,
     verifyApiKeyResourcePolicyAccess,
+    verifyApiKeyAiProviderAccess,
+    verifyApiKeyAiModelAccess,
     verifyUserHasAction
 } from "@server/middlewares";
 import HttpCode from "@server/types/HttpCode";
@@ -243,6 +246,26 @@ authenticated.get(
     siteResource.listSiteResourceClients
 );
 
+authenticated.get(
+    [
+        "/site-resource/:siteResourceId/ai-models",
+        "/private-resource/:siteResourceId/ai-models"
+    ],
+    verifyApiKeySiteResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.listResourceAiModels),
+    siteResource.listSiteResourceAiModels
+);
+
+authenticated.get(
+    [
+        "/site-resource/:siteResourceId/ai-providers",
+        "/private-resource/:siteResourceId/ai-providers"
+    ],
+    verifyApiKeySiteResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.listResourceAiModels),
+    siteResource.listSiteResourceAiProviders
+);
+
 authenticated.post(
     [
         "/site-resource/:siteResourceId/roles",
@@ -295,6 +318,72 @@ authenticated.post(
     siteResource.removeRoleFromSiteResource
 );
 
+authenticated.post(
+    [
+        "/site-resource/:siteResourceId/ai-models",
+        "/private-resource/:siteResourceId/ai-models"
+    ],
+    verifyApiKeySiteResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.setSiteResourceAiModels
+);
+
+authenticated.post(
+    [
+        "/site-resource/:siteResourceId/ai-models/add",
+        "/private-resource/:siteResourceId/ai-models/add"
+    ],
+    verifyApiKeySiteResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.addAiModelToSiteResource
+);
+
+authenticated.post(
+    [
+        "/site-resource/:siteResourceId/ai-models/remove",
+        "/private-resource/:siteResourceId/ai-models/remove"
+    ],
+    verifyApiKeySiteResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.removeAiModelFromSiteResource
+);
+
+authenticated.post(
+    [
+        "/site-resource/:siteResourceId/ai-providers",
+        "/private-resource/:siteResourceId/ai-providers"
+    ],
+    verifyApiKeySiteResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.setSiteResourceAiProviders
+);
+
+authenticated.post(
+    [
+        "/site-resource/:siteResourceId/ai-providers/add",
+        "/private-resource/:siteResourceId/ai-providers/add"
+    ],
+    verifyApiKeySiteResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.addAiProviderToSiteResource
+);
+
+authenticated.post(
+    [
+        "/site-resource/:siteResourceId/ai-providers/remove",
+        "/private-resource/:siteResourceId/ai-providers/remove"
+    ],
+    verifyApiKeySiteResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    siteResource.removeAiProviderFromSiteResource
+);
+
 authenticated.post(
     [
         "/site-resource/:siteResourceId/users/add",
@@ -507,6 +596,26 @@ authenticated.get(
     resource.listResourceUsers
 );
 
+authenticated.get(
+    [
+        "/resource/:resourceId/ai-models",
+        "/public-resource/:resourceId/ai-models"
+    ],
+    verifyApiKeyResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.listResourceAiModels),
+    resource.listResourceAiModels
+);
+
+authenticated.get(
+    [
+        "/resource/:resourceId/ai-providers",
+        "/public-resource/:resourceId/ai-providers"
+    ],
+    verifyApiKeyResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.listResourceAiModels),
+    resource.listResourceAiProviders
+);
+
 authenticated.get(
     ["/resource/:resourceId", "/public-resource/:resourceId"],
     verifyApiKeyResourceAccess,
@@ -708,6 +817,28 @@ authenticated.post(
     resource.setResourceRoles
 );
 
+authenticated.post(
+    [
+        "/resource/:resourceId/ai-models",
+        "/public-resource/:resourceId/ai-models"
+    ],
+    verifyApiKeyResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.setResourceAiModels
+);
+
+authenticated.post(
+    [
+        "/resource/:resourceId/ai-providers",
+        "/public-resource/:resourceId/ai-providers"
+    ],
+    verifyApiKeyResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.setResourceAiProviders
+);
+
 authenticated.post(
     ["/resource/:resourceId/users", "/public-resource/:resourceId/users"],
     verifyApiKeyResourceAccess,
@@ -900,6 +1031,50 @@ authenticated.post(
     resource.removeRoleFromResource
 );
 
+authenticated.post(
+    [
+        "/resource/:resourceId/ai-models/add",
+        "/public-resource/:resourceId/ai-models/add"
+    ],
+    verifyApiKeyResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.addAiModelToResource
+);
+
+authenticated.post(
+    [
+        "/resource/:resourceId/ai-models/remove",
+        "/public-resource/:resourceId/ai-models/remove"
+    ],
+    verifyApiKeyResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.removeAiModelFromResource
+);
+
+authenticated.post(
+    [
+        "/resource/:resourceId/ai-providers/add",
+        "/public-resource/:resourceId/ai-providers/add"
+    ],
+    verifyApiKeyResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.addAiProviderToResource
+);
+
+authenticated.post(
+    [
+        "/resource/:resourceId/ai-providers/remove",
+        "/public-resource/:resourceId/ai-providers/remove"
+    ],
+    verifyApiKeyResourceAccess,
+    verifyApiKeyHasAction(ActionsEnum.setResourceAiModels),
+    logActionAudit(ActionsEnum.setResourceAiModels),
+    resource.removeAiProviderFromResource
+);
+
 authenticated.post(
     [
         "/resource/:resourceId/users/add",
@@ -1366,3 +1541,95 @@ authenticated.get(
     verifyApiKeyHasAction(ActionsEnum.listResources),
     resource.listAllResourceNames
 );
+
+authenticated.put(
+    "/org/:orgId/ai-provider",
+    verifyApiKeyOrgAccess,
+    verifyApiKeyHasAction(ActionsEnum.createAiProvider),
+    logActionAudit(ActionsEnum.createAiProvider),
+    aiProvider.createAiProvider
+);
+
+authenticated.get(
+    "/org/:orgId/ai-providers",
+    verifyApiKeyOrgAccess,
+    verifyApiKeyHasAction(ActionsEnum.listAiProviders),
+    aiProvider.listAiProviders
+);
+
+authenticated.get(
+    "/ai-provider/:providerId",
+    verifyApiKeyAiProviderAccess,
+    verifyApiKeyHasAction(ActionsEnum.getAiProvider),
+    aiProvider.getAiProvider
+);
+
+authenticated.put(
+    "/ai-provider/:providerId/target",
+    verifyApiKeyAiProviderAccess,
+    verifyLimits,
+    verifyApiKeyHasAction(ActionsEnum.createTarget),
+    logActionAudit(ActionsEnum.createTarget),
+    target.createTarget
+);
+
+authenticated.get(
+    "/ai-provider/:providerId/targets",
+    verifyApiKeyAiProviderAccess,
+    verifyApiKeyHasAction(ActionsEnum.listTargets),
+    target.listTargets
+);
+
+authenticated.post(
+    "/ai-provider/:providerId",
+    verifyApiKeyAiProviderAccess,
+    verifyApiKeyHasAction(ActionsEnum.updateAiProvider),
+    logActionAudit(ActionsEnum.updateAiProvider),
+    aiProvider.updateAiProvider
+);
+
+authenticated.delete(
+    "/ai-provider/:providerId",
+    verifyApiKeyAiProviderAccess,
+    verifyApiKeyHasAction(ActionsEnum.deleteAiProvider),
+    logActionAudit(ActionsEnum.deleteAiProvider),
+    aiProvider.deleteAiProvider
+);
+
+authenticated.put(
+    "/ai-provider/:providerId/model",
+    verifyApiKeyAiProviderAccess,
+    verifyApiKeyHasAction(ActionsEnum.createAiModel),
+    logActionAudit(ActionsEnum.createAiModel),
+    aiProvider.createAiModel
+);
+
+authenticated.get(
+    "/ai-provider/:providerId/models",
+    verifyApiKeyAiProviderAccess,
+    verifyApiKeyHasAction(ActionsEnum.listAiModels),
+    aiProvider.listAiModels
+);
+
+authenticated.get(
+    "/ai-model/:modelId",
+    verifyApiKeyAiModelAccess,
+    verifyApiKeyHasAction(ActionsEnum.getAiModel),
+    aiProvider.getAiModel
+);
+
+authenticated.post(
+    "/ai-model/:modelId",
+    verifyApiKeyAiModelAccess,
+    verifyApiKeyHasAction(ActionsEnum.updateAiModel),
+    logActionAudit(ActionsEnum.updateAiModel),
+    aiProvider.updateAiModel
+);
+
+authenticated.delete(
+    "/ai-model/:modelId",
+    verifyApiKeyAiModelAccess,
+    verifyApiKeyHasAction(ActionsEnum.deleteAiModel),
+    logActionAudit(ActionsEnum.deleteAiModel),
+    aiProvider.deleteAiModel
+);

server/routers/internal.ts

@@ -63,3 +63,4 @@ internalRouter.use("/badger", badgerRouter);
 badgerRouter.post("/verify-session", badger.verifyResourceSession);
 
 badgerRouter.post("/exchange-session", badger.exchangeSession);
+

server/routers/launcher/formatLauncherAccess.ts

@@ -1,7 +1,8 @@
+import { SiteResource } from "@server/db";
 import { formatEndpoint, parseEndpoint } from "@server/lib/ip";
 
 export type SiteResourceDestinationInput = {
-    mode: "host" | "cidr" | "http" | "ssh";
+    mode: SiteResource["mode"];
     destination: string | null;
     destinationPort: number | null;
     scheme: "http" | "https" | null;

server/routers/newt/buildConfiguration.ts

@@ -15,7 +15,7 @@ import {
 } from "@server/db";
 import logger from "@server/logger";
 import { initPeerAddHandshake, updatePeer } from "../olm/peers";
-import { eq, and, inArray } from "drizzle-orm";
+import { eq, and, inArray, or, isNotNull, sql } from "drizzle-orm";
 import config from "@server/lib/config";
 import { decrypt } from "@server/lib/crypto";
 import {
@@ -211,7 +211,8 @@ export async function buildClientConfigurationForNewtClient(
     // call rather than letting each resource fetch its own — with thousands
     // of resources this avoids a concurrent DB/cache stampede for what is
     // often the very same (e.g. wildcard) certificate.
-    const certByDomain = await batchFetchCertsForSiteResources(allSiteResources);
+    const certByDomain =
+        await batchFetchCertsForSiteResources(allSiteResources);
 
     const resourceTargetsArr = await Promise.all(
         allSiteResources.map((resource) =>
@@ -240,7 +241,7 @@ export async function buildTargetConfigurationForNewtClient(
     version?: string | null,
     remoteExitNodeId?: string
 ) {
-    // Get all enabled targets with their resource mode information
+    // Get enabled HTTP/TCP/UDP targets for resources and AI providers
     const allTargets = await db
         .select({
             resourceId: targets.resourceId,
@@ -250,15 +251,18 @@ export async function buildTargetConfigurationForNewtClient(
             port: targets.port,
             internalPort: targets.internalPort,
             enabled: targets.enabled,
-            mode: resources.mode
+            mode: sql<string>`COALESCE(${resources.mode}, ${targets.mode})`.mapWith(
+                String
+            )
         })
         .from(targets)
-        .innerJoin(resources, eq(targets.resourceId, resources.resourceId))
+        .leftJoin(resources, eq(targets.resourceId, resources.resourceId))
         .where(
             and(
                 eq(targets.siteId, siteId),
                 eq(targets.enabled, true),
-                inArray(targets.mode, ["http", "udp", "tcp"])
+                inArray(targets.mode, ["http", "udp", "tcp"]),
+                or(isNotNull(targets.resourceId), isNotNull(targets.providerId))
             )
         );
 

server/routers/newt/handleNewtExitNodesRequestMessage.ts

@@ -2,14 +2,17 @@ import { db, sites } from "@server/db";
 import { MessageHandler } from "@server/routers/ws";
 import { exitNodes, Newt } from "@server/db";
 import logger from "@server/logger";
-import { ne, eq, or, and, count } from "drizzle-orm";
+import { eq } from "drizzle-orm";
 import { listExitNodes } from "#dynamic/lib/exitNodes";
+import { calculateExitNodeWeight } from "@server/lib/exitNodes";
 
-export const handleNewtPingRequestMessage: MessageHandler = async (context) => {
+export const handleNewtExitNodesRequestMessage: MessageHandler = async (
+    context
+) => {
     const { message, client, sendToClient } = context;
     const newt = client as Newt;
 
-    logger.info("Handling ping request newt message!");
+    logger.info("Handling exit nodes request newt message!");
 
     if (!newt) {
         logger.warn("Newt not found");
@@ -54,32 +57,13 @@ export const handleNewtPingRequestMessage: MessageHandler = async (context) => {
 
     const exitNodesPayload = await Promise.all(
         exitNodesList.map(async (node) => {
-            // (MAX_CONNECTIONS - current_connections) / MAX_CONNECTIONS)
-            // higher = more desirable
-            // like saying, this node has x% of its capacity left
+            const weight = await calculateExitNodeWeight(
+                node.exitNodeId,
+                node.maxConnections
+            );
 
-            let weight = 1;
-            const maxConnections = node.maxConnections;
-            if (maxConnections !== null && maxConnections !== undefined) {
-                const [currentConnections] = await db
-                    .select({
-                        count: count()
-                    })
-                    .from(sites)
-                    .where(
-                        and(
-                            eq(sites.exitNodeId, node.exitNodeId),
-                            eq(sites.online, true)
-                        )
-                    );
-
-                if (currentConnections.count >= maxConnections) {
-                    return null;
-                }
-
-                weight =
-                    (maxConnections - currentConnections.count) /
-                    maxConnections;
+            if (weight === null) {
+                return null;
             }
 
             return {

server/routers/newt/handleNewtGetConfigMessage.ts

@@ -95,16 +95,16 @@ export const handleNewtGetConfigMessage: MessageHandler = async (context) => {
             .limit(1);
         if (
             exitNode.reachableAt &&
-            existingSite.subnet &&
+            existingSite.exitNodeSubnet &&
             existingSite.listenPort
         ) {
             const payload = {
                 oldDestination: {
-                    destinationIP: existingSite.subnet?.split("/")[0],
+                    destinationIP: existingSite.exitNodeSubnet?.split("/")[0],
                     destinationPort: existingSite.listenPort || 1 // this satisfies gerbil for now but should be reevaluated
                 },
                 newDestination: {
-                    destinationIP: site.subnet?.split("/")[0],
+                    destinationIP: site.exitNodeSubnet?.split("/")[0],
                     destinationPort: site.listenPort || 1 // this satisfies gerbil for now but should be reevaluated
                 }
             };
@@ -132,7 +132,10 @@ export const handleNewtGetConfigMessage: MessageHandler = async (context) => {
         ({ targets: dedupedTargets, certs } = dedupeCertsForTargets(targets));
     }
 
-    const targetsToSend = await convertTargetsIfNecessary(newt.newtId, dedupedTargets); // for backward compatibility with old newt versions that don't support the new target format
+    const targetsToSend = await convertTargetsIfNecessary(
+        newt.newtId,
+        dedupedTargets
+    ); // for backward compatibility with old newt versions that don't support the new target format
 
     return {
         message: {

server/routers/newt/handleNewtRegisterMessage.ts

@@ -1,30 +1,20 @@
-import { db, ExitNode, newts, remoteExitNodes, Transaction } from "@server/db";
+import { db, newts, remoteExitNodes } from "@server/db";
 import { MessageHandler } from "@server/routers/ws";
 import { exitNodes, Newt, sites } from "@server/db";
 import { eq } from "drizzle-orm";
 import { addPeer, deletePeer } from "../gerbil/peers";
 import logger from "@server/logger";
 import config from "@server/lib/config";
-import { findNextAvailableCidr } from "@server/lib/ip";
 import {
+    ExitNodePingResult,
     selectBestExitNode,
     verifyExitNodeOrgAccess
 } from "#dynamic/lib/exitNodes";
+import { getUniqueSubnetForExitNode } from "@server/lib/exitNodes";
 import { fetchContainers } from "./dockerSocket";
-import { lockManager } from "#dynamic/lib/lock";
 import { buildTargetConfigurationForNewtClient } from "./buildConfiguration";
 import { canCompress } from "@server/lib/clientVersionChecks";
 
-export type ExitNodePingResult = {
-    exitNodeId: number;
-    latencyMs: number;
-    weight: number;
-    error?: string;
-    exitNodeName: string;
-    endpoint: string;
-    wasPreviouslyConnected: boolean;
-};
-
 export const handleNewtRegisterMessage: MessageHandler = async (context) => {
     const { message, client, sendToClient } = context;
     const newt = client as Newt;
@@ -94,9 +84,12 @@ export const handleNewtRegisterMessage: MessageHandler = async (context) => {
         fetchContainers(newt.newtId);
     }
 
-    let siteSubnet = oldSite.subnet;
+    let siteSubnet = oldSite.exitNodeSubnet;
     let exitNodeIdToQuery = oldSite.exitNodeId;
-    if (exitNodeId && (oldSite.exitNodeId !== exitNodeId || !oldSite.subnet)) {
+    if (
+        exitNodeId &&
+        (oldSite.exitNodeId !== exitNodeId || !oldSite.exitNodeSubnet)
+    ) {
         // This effectively moves the exit node to the new one
         exitNodeIdToQuery = exitNodeId; // Use the provided exitNodeId if it differs from the site's exitNodeId
 
@@ -115,7 +108,7 @@ export const handleNewtRegisterMessage: MessageHandler = async (context) => {
             return;
         }
 
-        const newSubnet = await getUniqueSubnetForSite(exitNode);
+        const newSubnet = await getUniqueSubnetForExitNode(exitNode);
 
         if (!newSubnet) {
             logger.error(
@@ -131,7 +124,7 @@ export const handleNewtRegisterMessage: MessageHandler = async (context) => {
             .set({
                 pubKey: publicKey,
                 exitNodeId: exitNodeId,
-                subnet: newSubnet
+                exitNodeSubnet: newSubnet
             })
             .where(eq(sites.siteId, siteId))
             .returning();
@@ -250,40 +243,3 @@ export const handleNewtRegisterMessage: MessageHandler = async (context) => {
         excludeSender: false // Include sender in broadcast
     };
 };
-
-async function getUniqueSubnetForSite(
-    exitNode: ExitNode,
-    trx: Transaction | typeof db = db
-): Promise<string | null> {
-    const lockKey = `subnet-allocation:${exitNode.exitNodeId}`;
-
-    return await lockManager.withLock(
-        lockKey,
-        async () => {
-            const sitesQuery = await trx
-                .select({
-                    subnet: sites.subnet
-                })
-                .from(sites)
-                .where(eq(sites.exitNodeId, exitNode.exitNodeId));
-
-            const blockSize = config.getRawConfig().gerbil.site_block_size;
-            const subnets = sitesQuery
-                .map((site) => site.subnet)
-                .filter(
-                    (subnet) =>
-                        subnet &&
-                        /^(\d{1,3}\.){3}\d{1,3}\/\d{1,2}$/.test(subnet)
-                )
-                .filter((subnet) => subnet !== null);
-            subnets.push(exitNode.address.replace(/\/\d+$/, `/${blockSize}`));
-            const newSubnet = findNextAvailableCidr(
-                subnets,
-                blockSize,
-                exitNode.address
-            );
-            return newSubnet;
-        },
-        5000 // 5 second lock TTL - subnet allocation should be quick
-    );
-}

server/routers/newt/index.ts

@@ -5,7 +5,7 @@ export * from "./handleNewtRegisterMessage";
 export * from "./handleReceiveBandwidthMessage";
 export * from "./handleNewtGetConfigMessage";
 export * from "./handleSocketMessages";
-export * from "./handleNewtPingRequestMessage";
+export * from "./handleNewtExitNodesRequestMessage";
 export * from "./handleApplyBlueprintMessage";
 export * from "./handleNewtPingMessage";
 export * from "./handleNewtDisconnectingMessage";

server/routers/olm/buildConfiguration.ts

@@ -19,6 +19,7 @@ import logger from "@server/logger";
 import { and, eq, inArray } from "drizzle-orm";
 import { addPeer, deletePeer } from "../newt/peers";
 import config from "@server/lib/config";
+import { SiR } from "react-icons/si";
 
 export async function buildSiteConfigurationForOlmClient(
     client: Client,
@@ -38,6 +39,8 @@ export async function buildSiteConfigurationForOlmClient(
         aliases: Alias[];
     }[] = [];
 
+    let exitNodeAliases: string[] = [];
+
     // Get all sites data
     const sitesData = await db
         .select()
@@ -48,10 +51,6 @@ export async function buildSiteConfigurationForOlmClient(
         )
         .where(eq(clientSitesAssociationsCache.clientId, client.clientId));
 
-    if (sitesData.length === 0) {
-        return siteConfigurations;
-    }
-
     // Batch-fetch every site resource this client has access to across ALL sites
     // in a single query, then group by siteId in memory. This avoids issuing one
     // query per site (which would be N round-trips for N sites).
@@ -68,8 +67,8 @@ export async function buildSiteConfigurationForOlmClient(
                 clientSiteResourcesAssociationsCache.siteResourceId
             )
         )
-        .innerJoin(networks, eq(siteResources.networkId, networks.networkId))
-        .innerJoin(siteNetworks, eq(networks.networkId, siteNetworks.networkId))
+        .leftJoin(networks, eq(siteResources.networkId, networks.networkId))
+        .leftJoin(siteNetworks, eq(networks.networkId, siteNetworks.networkId))
         .where(
             and(
                 eq(
@@ -81,7 +80,15 @@ export async function buildSiteConfigurationForOlmClient(
         );
 
     const siteResourcesBySiteId = new Map<number, SiteResource[]>();
+    let siteResourcesForExitNode = [];
     for (const row of allClientSiteResources) {
+        if (row.siteResource.requiresExitNodeConnection) {
+            siteResourcesForExitNode.push(row.siteResource);
+        }
+        if (!row.siteId) {
+            // because we are doing a leftJoin above to get the inference resources without a network / sites
+            continue;
+        }
         const arr = siteResourcesBySiteId.get(row.siteId);
         if (arr) {
             arr.push(row.siteResource);
@@ -90,6 +97,17 @@ export async function buildSiteConfigurationForOlmClient(
         }
     }
 
+    exitNodeAliases = siteResourcesForExitNode
+        .map((sr) => sr.fullDomain || sr.alias) // take either in case we introduce different resource types that don't have a fullDomain
+        .filter((a) => a != null);
+
+    if (sitesData.length == 0) {
+        return {
+            siteConfigurations,
+            exitNodeAliases
+        };
+    }
+
     // Batch-fetch exit nodes for all sites in one query (only needed in relay mode).
     const exitNodesById = new Map<number, typeof exitNodes.$inferSelect>();
     if (!jitMode && relay) {
@@ -167,7 +185,7 @@ export async function buildSiteConfigurationForOlmClient(
             peerOps.push(deletePeer(site.siteId, client.pubKey!));
         }
 
-        if (!site.subnet) {
+        if (!site.exitNodeSubnet) {
             logger.debug(`Site ${site.siteId} has no subnet, skipping`);
             continue;
         }
@@ -226,5 +244,8 @@ export async function buildSiteConfigurationForOlmClient(
         });
     }
 
-    return siteConfigurations;
+    return {
+        siteConfigurations,
+        exitNodeAliases
+    };
 }

server/routers/olm/handleOlmExitNodesRequestMessage.ts

@@ -0,0 +1,93 @@
+import { db, clients } from "@server/db";
+import { MessageHandler } from "@server/routers/ws";
+import { exitNodes, Olm } from "@server/db";
+import logger from "@server/logger";
+import { eq } from "drizzle-orm";
+import { listExitNodes } from "#dynamic/lib/exitNodes";
+import { calculateExitNodeWeight } from "@server/lib/exitNodes";
+
+export const handleOlmExitNodesRequestMessage: MessageHandler = async (
+    context
+) => {
+    const { message, client: olmClient, sendToClient } = context;
+    const olm = olmClient as Olm;
+
+    logger.info("Handling exit nodes request olm message!");
+
+    if (!olm) {
+        logger.warn("olm not found");
+        return;
+    }
+
+    // Get the olm's orgId through the client relationship
+    if (!olm.clientId) {
+        logger.warn("olm clientId not found");
+        return;
+    }
+
+    const [client] = await db
+        .select({ orgId: clients.orgId })
+        .from(clients)
+        .where(eq(clients.clientId, olm.clientId))
+        .limit(1);
+
+    if (!client || !client.orgId) {
+        logger.warn("client not found");
+        return;
+    }
+
+    const { noCloud, chainId } = message.data;
+
+    const exitNodesList = await listExitNodes(
+        client.orgId,
+        true,
+        noCloud || false,
+        olm.clientId
+    ); // filter for only the online ones
+
+    let lastExitNodeId = null;
+    if (olm.clientId) {
+        const [lastExitNode] = await db
+            .select()
+            .from(clients)
+            .where(eq(clients.clientId, olm.clientId))
+            .limit(1);
+        lastExitNodeId = lastExitNode?.exitNodeId || null;
+    }
+
+    const exitNodesPayload = await Promise.all(
+        exitNodesList.map(async (node) => {
+            const weight = await calculateExitNodeWeight(
+                node.exitNodeId,
+                node.maxConnections
+            );
+
+            if (weight === null) {
+                return null;
+            }
+
+            return {
+                exitNodeId: node.exitNodeId,
+                exitNodeName: node.name,
+                endpoint: node.endpoint,
+                weight,
+                wasPreviouslyConnected: node.exitNodeId === lastExitNodeId
+            };
+        })
+    );
+
+    // filter out null values
+    const filteredExitNodes = exitNodesPayload.filter((node) => node !== null);
+
+    return {
+        message: {
+            type: "olm/ping/exitNodes",
+            data: {
+                exitNodes: filteredExitNodes,
+                chainId: chainId
+            }
+        },
+        broadcast: false, // Send to all clients
+        excludeSender: false // Include sender in broadcast
+    };
+};