feat(mcpd): inference proxy — POST /api/v1/llms/:name/infer

Why: the point of the Llm resource (Phase 1) is that credentials never leave
the server. This lands the proxy: clients POST OpenAI chat/completions to
mcpd, mcpd attaches the provider API key server-side, and the response
streams back as OpenAI-format SSE.

Design:
- Wire format client-side is always OpenAI chat/completions — every existing
  SDK speaks it. Adapters translate on the provider side.
- `openai | vllm | deepseek | ollama` → pure passthrough (they already speak
  OpenAI). `anthropic` → translator to/from Anthropic Messages API
  (system-string extraction, content-block flattening, SSE event remap).
- Plain fetch; no @anthropic-ai/sdk dep. Consistent with the OpenBao driver
  shape and keeps the proxy layer thin.
- `gemini-cli` intentionally rejected — subprocess providers need extra
  lifecycle plumbing; deferred to a follow-up.
- Streaming: adapters yield `StreamingChunk`s; the route frames them as
  `data: <json>\n\n` + terminal `data: [DONE]\n\n` so any OpenAI client
  works unchanged.

RBAC:
- New URL special-case in mapUrlToPermission: `POST /api/v1/llms/:name/infer`
  → `run:llms:<name>` (not the default create:llms). Users need an explicit
  `{role: 'run', resource: 'llms', [name: X]}` binding to call infer.
- Possession of `edit:llms` does NOT imply `run` — keeps catalogue
  management separate from spend.

Audit: route emits an `llm_inference_call` event per request (llm name,
model, user/tokenSha, streaming, duration, status). main.ts wires it to the
structured logger for now; hook is in place for a richer audit sink later.

Tests:
- 11 adapter tests (passthrough POST shape + default URLs + no-auth ollama +
  SSE forwarding; anthropic translate request/response + non-2xx wrap + SSE
  event translation; registry dispatch + caching + unsupported-provider).
- 7 route tests (404, 400, non-streaming dispatch + audit, apiKey failure,
  null apiKeyRef path, streaming SSE output, 502 on adapter error).
- Full suite 1830/1830 (+18 from Phase 1's 1812). TypeScript clean.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/mcpd/src/main.ts

@@ -28,7 +28,9 @@ import { registerSecretBackendRoutes } from './routes/secret-backends.js';
 import { registerSecretMigrateRoutes } from './routes/secret-migrate.js';
 import { LlmRepository } from './repositories/llm.repository.js';
 import { LlmService } from './services/llm.service.js';
+import { LlmAdapterRegistry } from './services/llm/dispatcher.js';
 import { registerLlmRoutes } from './routes/llms.js';
+import { registerLlmInferRoutes } from './routes/llm-infer.js';
 import { PromptRepository } from './repositories/prompt.repository.js';
 import { PromptRequestRepository } from './repositories/prompt-request.repository.js';
 import { bootstrapSystemProject } from './bootstrap/system-project.js';
@@ -105,6 +107,12 @@ function mapUrlToPermission(method: string, url: string): PermissionCheck {
   // /api/v1/secrets/migrate is a bulk cross-backend operation — treat as op, not a plain secret write.
   if (url.startsWith('/api/v1/secrets/migrate')) return { kind: 'operation', operation: 'migrate-secrets' };
 
+  // /api/v1/llms/:name/infer → `run:llms:<name>` (not the default create:llms).
+  const inferMatch = url.match(/^\/api\/v1\/llms\/([^/?]+)\/infer/);
+  if (inferMatch?.[1]) {
+    return { kind: 'resource', resource: 'llms', action: 'run', resourceName: inferMatch[1] };
+  }
+
   const resourceMap: Record<string, string | undefined> = {
     'servers': 'servers',
     'instances': 'instances',
@@ -334,6 +342,7 @@ async function main(): Promise<void> {
   const secretService = new SecretService(secretRepo, secretBackendService);
   const secretMigrateService = new SecretMigrateService(secretRepo, secretBackendService);
   const llmService = new LlmService(llmRepo, secretService);
+  const llmAdapters = new LlmAdapterRegistry();
   const instanceService = new InstanceService(instanceRepo, serverRepo, orchestrator, secretService);
   serverService.setInstanceService(instanceService);
   const projectService = new ProjectService(projectRepo, serverRepo);
@@ -475,6 +484,23 @@ async function main(): Promise<void> {
   registerSecretBackendRoutes(app, secretBackendService);
   registerSecretMigrateRoutes(app, secretMigrateService);
   registerLlmRoutes(app, llmService);
+  registerLlmInferRoutes(app, {
+    llmService,
+    adapters: llmAdapters,
+    onInferenceEvent: (event) => {
+      app.log.info({
+        event: 'llm_inference_call',
+        llm: event.llmName,
+        model: event.model,
+        type: event.type,
+        userId: event.userId,
+        tokenSha: event.tokenSha,
+        streaming: event.streaming,
+        durationMs: event.durationMs,
+        status: event.status,
+      });
+    },
+  });
   registerInstanceRoutes(app, instanceService);
   registerProjectRoutes(app, projectService);
   registerAuditLogRoutes(app, auditLogService);

src/mcpd/src/routes/llm-infer.ts

@@ -0,0 +1,145 @@
+/**
+ * POST /api/v1/llms/:name/infer
+ *
+ * OpenAI-compatible chat completions endpoint. The RBAC check runs in the
+ * global hook — this URL maps to `run:llms:<name>`, not the default
+ * `create:llms`. See `main.ts:mapUrlToPermission`.
+ *
+ * Non-streaming: resolves the Llm, dispatches to the right provider adapter,
+ * returns the OpenAI chat.completion JSON.
+ *
+ * Streaming (`stream: true`): pipes adapter-emitted chunks back as
+ * `text/event-stream`. Adapters translate provider-native SSE into OpenAI
+ * `chat.completion.chunk`s so clients can use any OpenAI SDK unchanged.
+ */
+import type { FastifyInstance, FastifyReply } from 'fastify';
+import type { LlmService } from '../services/llm.service.js';
+import type { LlmAdapterRegistry } from '../services/llm/dispatcher.js';
+import { NotFoundError } from '../services/mcp-server.service.js';
+import type { OpenAiChatRequest, InferContext } from '../services/llm/types.js';
+
+export interface LlmInferDeps {
+  llmService: LlmService;
+  adapters: LlmAdapterRegistry;
+  /** Optional hook to emit audit events — consumer may ignore. */
+  onInferenceEvent?: (event: InferenceAuditEvent) => void;
+}
+
+export interface InferenceAuditEvent {
+  kind: 'llm_inference_call';
+  llmName: string;
+  model: string;
+  type: string;
+  userId?: string | undefined;
+  tokenSha?: string | undefined;
+  streaming: boolean;
+  durationMs: number;
+  status: number;
+}
+
+export function registerLlmInferRoutes(
+  app: FastifyInstance,
+  deps: LlmInferDeps,
+): void {
+  app.post<{ Params: { name: string }; Body: OpenAiChatRequest }>(
+    '/api/v1/llms/:name/infer',
+    async (request, reply) => {
+      const started = Date.now();
+      let llm;
+      try {
+        llm = await deps.llmService.getByName(request.params.name);
+      } catch (err) {
+        if (err instanceof NotFoundError) {
+          reply.code(404);
+          return { error: err.message };
+        }
+        throw err;
+      }
+
+      const body = (request.body ?? {}) as OpenAiChatRequest;
+      if (!body.messages || body.messages.length === 0) {
+        reply.code(400);
+        return { error: 'messages is required' };
+      }
+
+      // Resolve API key (may be empty string for providers that don't take one).
+      let apiKey = '';
+      if (llm.apiKeyRef !== null) {
+        try {
+          apiKey = await deps.llmService.resolveApiKey(llm.name);
+        } catch (err) {
+          reply.code(500);
+          return { error: `Failed to resolve API key: ${err instanceof Error ? err.message : String(err)}` };
+        }
+      }
+
+      const ctx: InferContext = {
+        body,
+        modelOverride: llm.model,
+        apiKey,
+        url: llm.url,
+        extraConfig: llm.extraConfig,
+      };
+
+      const adapter = deps.adapters.get(llm.type);
+      const streaming = body.stream === true;
+
+      const audit = (status: number): void => {
+        if (deps.onInferenceEvent === undefined) return;
+        deps.onInferenceEvent({
+          kind: 'llm_inference_call',
+          llmName: llm.name,
+          model: llm.model,
+          type: llm.type,
+          userId: request.userId,
+          tokenSha: request.mcpToken?.tokenSha,
+          streaming,
+          durationMs: Date.now() - started,
+          status,
+        });
+      };
+
+      if (!streaming) {
+        try {
+          const result = await adapter.infer(ctx);
+          reply.code(result.status);
+          audit(result.status);
+          return result.body;
+        } catch (err) {
+          audit(502);
+          reply.code(502);
+          return { error: err instanceof Error ? err.message : String(err) };
+        }
+      }
+
+      // Streaming path — set SSE headers and pipe chunks.
+      reply.raw.writeHead(200, {
+        'Content-Type': 'text/event-stream',
+        'Cache-Control': 'no-cache',
+        Connection: 'keep-alive',
+        'X-Accel-Buffering': 'no',
+      });
+      try {
+        for await (const chunk of adapter.stream(ctx)) {
+          writeSseChunk(reply, chunk.data);
+          if (chunk.done === true) break;
+        }
+        audit(200);
+      } catch (err) {
+        const payload = JSON.stringify({
+          error: { message: err instanceof Error ? err.message : String(err) },
+        });
+        writeSseChunk(reply, payload);
+        writeSseChunk(reply, '[DONE]');
+        audit(502);
+      } finally {
+        reply.raw.end();
+      }
+      return reply;
+    },
+  );
+}
+
+function writeSseChunk(reply: FastifyReply, data: string): void {
+  reply.raw.write(`data: ${data}\n\n`);
+}

src/mcpd/src/services/llm/adapters/anthropic.ts

@@ -0,0 +1,256 @@
+/**
+ * Anthropic adapter — translates between OpenAI chat/completions format and
+ * the Anthropic Messages API (`POST /v1/messages`).
+ *
+ * Key differences we translate:
+ *   - OpenAI `role: 'system'` messages become a top-level `system` string.
+ *   - Anthropic returns `content: [{ type: 'text', text }]` — we join into
+ *     OpenAI's `content: "…"` string.
+ *   - Streaming: Anthropic emits a sequence of
+ *     `message_start / content_block_{start,delta,stop} / message_delta /
+ *     message_stop` events. We translate those to OpenAI
+ *     `chat.completion.chunk` deltas.
+ *
+ * This adapter implements the subset needed for plain-text chat — tool-use
+ * translation is intentionally left out for this phase; agents that need tool
+ * calling should target an OpenAI-compatible provider until the translator
+ * covers it.
+ */
+import type {
+  LlmAdapter,
+  InferContext,
+  NonStreamingResult,
+  StreamingChunk,
+  AdapterDeps,
+  OpenAiMessage,
+} from '../types.js';
+
+const DEFAULT_ANTHROPIC_URL = 'https://api.anthropic.com';
+const ANTHROPIC_VERSION = '2023-06-01';
+
+interface AnthropicMessageResponse {
+  id: string;
+  model: string;
+  role: 'assistant';
+  content: Array<{ type: 'text'; text: string } | { type: string; [k: string]: unknown }>;
+  stop_reason?: string;
+  usage?: { input_tokens: number; output_tokens: number };
+}
+
+export class AnthropicAdapter implements LlmAdapter {
+  readonly kind = 'anthropic';
+  private readonly fetchImpl: typeof globalThis.fetch;
+
+  constructor(deps: AdapterDeps = {}) {
+    this.fetchImpl = deps.fetch ?? globalThis.fetch;
+  }
+
+  async infer(ctx: InferContext): Promise<NonStreamingResult> {
+    const url = (ctx.url !== '' ? ctx.url : DEFAULT_ANTHROPIC_URL).replace(/\/+$/, '');
+    const body = this.toAnthropicRequest(ctx, false);
+    const res = await this.fetchImpl(`${url}/v1/messages`, {
+      method: 'POST',
+      headers: this.headers(ctx),
+      body: JSON.stringify(body),
+    });
+    if (!res.ok) {
+      const text = await res.text().catch(() => '');
+      return {
+        status: res.status,
+        body: { error: { message: `anthropic: HTTP ${String(res.status)} ${text}` } },
+      };
+    }
+    const anth = await res.json() as AnthropicMessageResponse;
+    return { status: 200, body: this.toOpenAiResponse(anth) };
+  }
+
+  async *stream(ctx: InferContext): AsyncGenerator<StreamingChunk> {
+    const url = (ctx.url !== '' ? ctx.url : DEFAULT_ANTHROPIC_URL).replace(/\/+$/, '');
+    const body = this.toAnthropicRequest(ctx, true);
+    const res = await this.fetchImpl(`${url}/v1/messages`, {
+      method: 'POST',
+      headers: this.headers(ctx),
+      body: JSON.stringify(body),
+    });
+    if (!res.ok || res.body === null) {
+      const text = await res.text().catch(() => '');
+      throw new Error(`anthropic stream: HTTP ${String(res.status)} ${text}`);
+    }
+
+    const id = `chatcmpl-${cryptoNonce()}`;
+    const model = body.model;
+    const created = Math.floor(Date.now() / 1000);
+
+    // Parse Anthropic SSE. Each event is `event: <name>\ndata: <json>\n\n`.
+    const decoder = new TextDecoder();
+    let buf = '';
+    const reader = res.body.getReader();
+    let emittedFirst = false;
+
+    const baseChunk = (delta: Record<string, unknown>, finishReason?: string): string => {
+      const chunk = {
+        id,
+        object: 'chat.completion.chunk',
+        created,
+        model,
+        choices: [{
+          index: 0,
+          delta,
+          finish_reason: finishReason ?? null,
+        }],
+      };
+      return JSON.stringify(chunk);
+    };
+
+    try {
+      // eslint-disable-next-line no-constant-condition
+      while (true) {
+        const { value, done } = await reader.read();
+        if (done) break;
+        buf += decoder.decode(value, { stream: true });
+
+        let idx: number;
+        while ((idx = buf.indexOf('\n\n')) !== -1) {
+          const rawEvent = buf.slice(0, idx);
+          buf = buf.slice(idx + 2);
+          const parsed = parseSseEvent(rawEvent);
+          if (parsed === null) continue;
+          const { event, data } = parsed;
+
+          if (event === 'content_block_delta') {
+            const textDelta = (data as { delta?: { type?: string; text?: string } }).delta;
+            if (textDelta?.type === 'text_delta' && typeof textDelta.text === 'string') {
+              if (!emittedFirst) {
+                yield { data: baseChunk({ role: 'assistant', content: '' }) };
+                emittedFirst = true;
+              }
+              yield { data: baseChunk({ content: textDelta.text }) };
+            }
+          } else if (event === 'message_delta') {
+            const stopReason = (data as { delta?: { stop_reason?: string } }).delta?.stop_reason;
+            if (typeof stopReason === 'string') {
+              yield { data: baseChunk({}, mapStopReason(stopReason)) };
+            }
+          } else if (event === 'message_stop') {
+            yield { data: '[DONE]', done: true };
+            return;
+          } else if (event === 'error') {
+            throw new Error(`anthropic stream error: ${JSON.stringify(data)}`);
+          }
+        }
+      }
+    } finally {
+      reader.releaseLock();
+    }
+    // Anthropic closed without message_stop — give consumer a clean end.
+    yield { data: '[DONE]', done: true };
+  }
+
+  private headers(ctx: InferContext): Record<string, string> {
+    return {
+      'Content-Type': 'application/json',
+      'x-api-key': ctx.apiKey,
+      'anthropic-version': ANTHROPIC_VERSION,
+    };
+  }
+
+  /** Translate the OpenAI request to the Anthropic Messages shape. */
+  private toAnthropicRequest(ctx: InferContext, stream: boolean): {
+    model: string;
+    max_tokens: number;
+    messages: Array<{ role: 'user' | 'assistant'; content: string }>;
+    system?: string;
+    stream?: boolean;
+    temperature?: number;
+    top_p?: number;
+    stop_sequences?: string[];
+  } {
+    const { body } = ctx;
+    const systemParts: string[] = [];
+    const messages: Array<{ role: 'user' | 'assistant'; content: string }> = [];
+
+    for (const msg of body.messages) {
+      const text = normaliseContent(msg);
+      if (msg.role === 'system') {
+        systemParts.push(text);
+      } else if (msg.role === 'user' || msg.role === 'assistant') {
+        messages.push({ role: msg.role, content: text });
+      }
+      // `tool` role messages are dropped — tool translation is out of scope
+      // for this phase.
+    }
+
+    const out: ReturnType<typeof this.toAnthropicRequest> = {
+      model: body.model !== '' ? body.model : ctx.modelOverride,
+      max_tokens: typeof body.max_tokens === 'number' ? body.max_tokens : 1024,
+      messages,
+    };
+    if (systemParts.length > 0) out.system = systemParts.join('\n\n');
+    if (stream) out.stream = true;
+    if (typeof body.temperature === 'number') out.temperature = body.temperature;
+    if (typeof body.top_p === 'number') out.top_p = body.top_p;
+    if (body.stop !== undefined) {
+      out.stop_sequences = Array.isArray(body.stop) ? body.stop : [body.stop];
+    }
+    return out;
+  }
+
+  private toOpenAiResponse(anth: AnthropicMessageResponse): Record<string, unknown> {
+    const text = anth.content
+      .map((c) => (c.type === 'text' && typeof (c as { text?: unknown }).text === 'string'
+        ? (c as { text: string }).text
+        : ''))
+      .join('');
+    return {
+      id: `chatcmpl-${anth.id}`,
+      object: 'chat.completion',
+      created: Math.floor(Date.now() / 1000),
+      model: anth.model,
+      choices: [{
+        index: 0,
+        message: { role: 'assistant', content: text },
+        finish_reason: mapStopReason(anth.stop_reason ?? 'end_turn'),
+      }],
+      usage: anth.usage ? {
+        prompt_tokens: anth.usage.input_tokens,
+        completion_tokens: anth.usage.output_tokens,
+        total_tokens: anth.usage.input_tokens + anth.usage.output_tokens,
+      } : undefined,
+    };
+  }
+}
+
+function normaliseContent(msg: OpenAiMessage): string {
+  if (typeof msg.content === 'string') return msg.content;
+  return msg.content
+    .map((part) => (typeof part.text === 'string' ? part.text : ''))
+    .join('');
+}
+
+function mapStopReason(r: string): string {
+  // Anthropic → OpenAI finish_reason
+  if (r === 'end_turn' || r === 'stop_sequence') return 'stop';
+  if (r === 'max_tokens') return 'length';
+  if (r === 'tool_use') return 'tool_calls';
+  return r;
+}
+
+function parseSseEvent(raw: string): { event: string; data: unknown } | null {
+  let event = '';
+  let dataLine = '';
+  for (const line of raw.split('\n')) {
+    if (line.startsWith('event:')) event = line.slice(6).trim();
+    else if (line.startsWith('data:')) dataLine += line.slice(5).trim();
+  }
+  if (dataLine === '') return null;
+  try {
+    return { event, data: JSON.parse(dataLine) as unknown };
+  } catch {
+    return null;
+  }
+}
+
+function cryptoNonce(): string {
+  // Not security-sensitive — just a short randomish id.
+  return Math.random().toString(36).slice(2, 10);
+}

src/mcpd/src/services/llm/adapters/openai-passthrough.ts

@@ -0,0 +1,112 @@
+/**
+ * OpenAI-passthrough adapter.
+ *
+ * Covers any provider that already speaks OpenAI chat/completions on the
+ * wire: `openai`, `vllm`, `deepseek`, `ollama` (with their openai-compatible
+ * endpoint enabled). The adapter forwards the request body verbatim and
+ * streams the response straight through — no wire translation.
+ *
+ * Defaults when `url` is empty:
+ *   - openai → https://api.openai.com
+ *   - deepseek → https://api.deepseek.com
+ *   - vllm/ollama → must be configured; these have no canonical public URL.
+ */
+import type { LlmAdapter, InferContext, NonStreamingResult, StreamingChunk, AdapterDeps } from '../types.js';
+
+const DEFAULT_URLS: Record<string, string> = {
+  openai: 'https://api.openai.com',
+  deepseek: 'https://api.deepseek.com',
+};
+
+export class OpenAiPassthroughAdapter implements LlmAdapter {
+  readonly kind: string;
+  private readonly fetchImpl: typeof globalThis.fetch;
+
+  constructor(kind: 'openai' | 'vllm' | 'deepseek' | 'ollama', deps: AdapterDeps = {}) {
+    this.kind = kind;
+    this.fetchImpl = deps.fetch ?? globalThis.fetch;
+  }
+
+  async infer(ctx: InferContext): Promise<NonStreamingResult> {
+    const url = this.endpointUrl(ctx.url);
+    const body = this.prepareBody(ctx, false);
+    const res = await this.fetchImpl(`${url}/v1/chat/completions`, {
+      method: 'POST',
+      headers: this.headers(ctx),
+      body: JSON.stringify(body),
+    });
+    const json = await res.json() as unknown;
+    return { status: res.status, body: json };
+  }
+
+  async *stream(ctx: InferContext): AsyncGenerator<StreamingChunk> {
+    const url = this.endpointUrl(ctx.url);
+    const body = this.prepareBody(ctx, true);
+    const res = await this.fetchImpl(`${url}/v1/chat/completions`, {
+      method: 'POST',
+      headers: this.headers(ctx),
+      body: JSON.stringify(body),
+    });
+    if (!res.ok || res.body === null) {
+      const text = await res.text().catch(() => '');
+      throw new Error(`${this.kind} stream: HTTP ${String(res.status)} ${text}`);
+    }
+
+    // Re-frame the provider's SSE stream into our `StreamingChunk` shape.
+    // OpenAI-compat providers already emit `data: {...}` + `data: [DONE]` —
+    // we just unwrap the `data: ` prefix, forward payloads, and emit a
+    // single terminal `done` chunk so the consumer always gets one.
+    const decoder = new TextDecoder();
+    let buf = '';
+    const reader = res.body.getReader();
+    try {
+      // eslint-disable-next-line no-constant-condition
+      while (true) {
+        const { value, done } = await reader.read();
+        if (done) break;
+        buf += decoder.decode(value, { stream: true });
+        let idx: number;
+        while ((idx = buf.indexOf('\n\n')) !== -1) {
+          const event = buf.slice(0, idx);
+          buf = buf.slice(idx + 2);
+          for (const line of event.split('\n')) {
+            if (!line.startsWith('data:')) continue;
+            const payload = line.slice(5).trim();
+            if (payload === '') continue;
+            if (payload === '[DONE]') {
+              yield { data: '[DONE]', done: true };
+              return;
+            }
+            yield { data: payload };
+          }
+        }
+      }
+    } finally {
+      reader.releaseLock();
+    }
+    // Provider closed without emitting [DONE] — give the consumer a clean end.
+    yield { data: '[DONE]', done: true };
+  }
+
+  private endpointUrl(url: string): string {
+    if (url !== '') return url.replace(/\/+$/, '');
+    const def = DEFAULT_URLS[this.kind];
+    if (def === undefined) {
+      throw new Error(`${this.kind}: url is required (no default endpoint for this provider)`);
+    }
+    return def;
+  }
+
+  private headers(ctx: InferContext): Record<string, string> {
+    const headers: Record<string, string> = { 'Content-Type': 'application/json' };
+    if (ctx.apiKey !== '') headers['Authorization'] = `Bearer ${ctx.apiKey}`;
+    return headers;
+  }
+
+  private prepareBody(ctx: InferContext, stream: boolean): Record<string, unknown> {
+    const out: Record<string, unknown> = { ...ctx.body };
+    if (out.model === undefined || out.model === '') out.model = ctx.modelOverride;
+    out.stream = stream;
+    return out;
+  }
+}

src/mcpd/src/services/llm/dispatcher.ts

@@ -0,0 +1,52 @@
+/**
+ * Adapter dispatcher for the inference proxy.
+ *
+ * `getAdapter(type)` returns the right adapter instance for an Llm's `type`
+ * column. Adapters are cached per-type — they carry no per-request state.
+ * The caller (the infer route) supplies the resolved API key + request body
+ * through `InferContext`, so a single adapter instance serves every Llm of
+ * that type.
+ */
+import type { LlmAdapter, AdapterDeps } from './types.js';
+import { OpenAiPassthroughAdapter } from './adapters/openai-passthrough.js';
+import { AnthropicAdapter } from './adapters/anthropic.js';
+
+export class UnsupportedProviderError extends Error {
+  constructor(type: string) {
+    super(`Unsupported LLM provider: ${type}`);
+    this.name = 'UnsupportedProviderError';
+  }
+}
+
+export class LlmAdapterRegistry {
+  private readonly cache = new Map<string, LlmAdapter>();
+
+  constructor(private readonly deps: AdapterDeps = {}) {}
+
+  get(type: string): LlmAdapter {
+    const cached = this.cache.get(type);
+    if (cached !== undefined) return cached;
+    const adapter = this.build(type);
+    this.cache.set(type, adapter);
+    return adapter;
+  }
+
+  private build(type: string): LlmAdapter {
+    switch (type) {
+      case 'openai':
+      case 'vllm':
+      case 'deepseek':
+      case 'ollama':
+        return new OpenAiPassthroughAdapter(type, this.deps);
+      case 'anthropic':
+        return new AnthropicAdapter(this.deps);
+      case 'gemini-cli':
+        // Intentionally deferred — gemini-cli requires the binary on the mcpd
+        // pod filesystem and subprocess lifecycle management. Flagged as
+        // homelab-only in the plan; not landing in this phase.
+        throw new UnsupportedProviderError(`${type} (subprocess providers are not supported in the proxy yet)`);
+      default:
+        throw new UnsupportedProviderError(type);
+    }
+  }
+}

src/mcpd/src/services/llm/types.ts

@@ -0,0 +1,70 @@
+/**
+ * Shared types for the LLM inference proxy.
+ *
+ * The wire format on the mcpctl side is OpenAI's chat/completions v1 — it's
+ * the de-facto lingua franca and every client library already speaks it.
+ * Provider-specific adapters translate to/from that shape.
+ */
+
+export interface OpenAiMessage {
+  role: 'system' | 'user' | 'assistant' | 'tool';
+  content: string | Array<{ type: string; text?: string; [k: string]: unknown }>;
+  name?: string;
+  tool_call_id?: string;
+  tool_calls?: Array<{ id: string; type: 'function'; function: { name: string; arguments: string } }>;
+}
+
+export interface OpenAiChatRequest {
+  model: string;
+  messages: OpenAiMessage[];
+  stream?: boolean;
+  temperature?: number;
+  max_tokens?: number;
+  top_p?: number;
+  stop?: string | string[];
+  tools?: Array<{ type: 'function'; function: { name: string; description?: string; parameters?: Record<string, unknown> } }>;
+  tool_choice?: unknown;
+  // Passthrough: unknown extras forwarded as-is.
+  [k: string]: unknown;
+}
+
+export interface InferContext {
+  /** Normalised OpenAI-format body. Adapters read/transform from here. */
+  body: OpenAiChatRequest;
+  /** The Llm row's `model` field, used when the request body has an empty model. */
+  modelOverride: string;
+  /** The resolved API key, or empty string for providers that don't take one. */
+  apiKey: string;
+  /** Target URL from the Llm row (may be empty for provider-default). */
+  url: string;
+  /** Arbitrary config from the Llm row (e.g. vllm gpu settings). */
+  extraConfig: Record<string, unknown>;
+}
+
+export interface NonStreamingResult {
+  status: number;
+  /** OpenAI chat.completion response body. */
+  body: unknown;
+}
+
+export interface StreamingChunk {
+  /** Raw SSE data payload. Consumer emits `data: <payload>\n\n`. */
+  data: string;
+  /** Mark the end of stream — consumer emits `data: [DONE]\n\n`. */
+  done?: boolean;
+}
+
+export interface LlmAdapter {
+  readonly kind: string;
+  /** Non-streaming request. Returns the final chat.completion body. */
+  infer(ctx: InferContext): Promise<NonStreamingResult>;
+  /**
+   * Streaming request. Yields OpenAI-format SSE chunks. Adapters translate
+   * provider-native stream formats into OpenAI `chat.completion.chunk`s.
+   */
+  stream(ctx: InferContext): AsyncGenerator<StreamingChunk>;
+}
+
+export interface AdapterDeps {
+  fetch?: typeof globalThis.fetch;
+}

src/mcpd/tests/llm-adapters.test.ts

@@ -0,0 +1,210 @@
+import { describe, it, expect, vi } from 'vitest';
+import { OpenAiPassthroughAdapter } from '../src/services/llm/adapters/openai-passthrough.js';
+import { AnthropicAdapter } from '../src/services/llm/adapters/anthropic.js';
+import { LlmAdapterRegistry, UnsupportedProviderError } from '../src/services/llm/dispatcher.js';
+import type { InferContext } from '../src/services/llm/types.js';
+
+function mockFetch(responses: Array<{ match: RegExp; status: number; body?: unknown; text?: string }>): ReturnType<typeof vi.fn> {
+  return vi.fn(async (input: string | URL, _init?: RequestInit) => {
+    const url = String(input);
+    const match = responses.find((r) => r.match.test(url));
+    if (!match) throw new Error(`unexpected fetch: ${url}`);
+    const body = match.body !== undefined ? JSON.stringify(match.body) : (match.text ?? '');
+    return new Response(body, { status: match.status, headers: { 'Content-Type': 'application/json' } });
+  });
+}
+
+function makeCtx(overrides: Partial<InferContext> = {}): InferContext {
+  return {
+    body: { model: '', messages: [{ role: 'user', content: 'hello' }] },
+    modelOverride: 'default-model',
+    apiKey: 'test-key',
+    url: '',
+    extraConfig: {},
+    ...overrides,
+  };
+}
+
+// Helper to build a streaming Response from SSE lines.
+function sseResponse(events: string[]): Response {
+  const body = events.join('\n\n') + '\n\n';
+  const stream = new ReadableStream<Uint8Array>({
+    start(controller) {
+      controller.enqueue(new TextEncoder().encode(body));
+      controller.close();
+    },
+  });
+  return new Response(stream, { status: 200, headers: { 'Content-Type': 'text/event-stream' } });
+}
+
+describe('OpenAiPassthroughAdapter', () => {
+  it('infer: POSTs to <url>/v1/chat/completions with Authorization + body', async () => {
+    const fetchFn = mockFetch([{
+      match: /\/v1\/chat\/completions$/,
+      status: 200,
+      body: { id: 'x', choices: [{ message: { role: 'assistant', content: 'hi' } }] },
+    }]);
+    const adapter = new OpenAiPassthroughAdapter('openai', { fetch: fetchFn as unknown as typeof fetch });
+    const ctx = makeCtx({ url: 'https://api.example.com' });
+    const res = await adapter.infer(ctx);
+    expect(res.status).toBe(200);
+    const [url, init] = fetchFn.mock.calls[0] as [string, RequestInit];
+    expect(url).toBe('https://api.example.com/v1/chat/completions');
+    expect(init.method).toBe('POST');
+    const headers = init.headers as Record<string, string>;
+    expect(headers['Authorization']).toBe('Bearer test-key');
+    const sent = JSON.parse(init.body as string) as { model: string; stream: boolean };
+    expect(sent.model).toBe('default-model');  // filled from modelOverride
+    expect(sent.stream).toBe(false);
+  });
+
+  it('infer: uses default URL for openai when url is empty', async () => {
+    const fetchFn = mockFetch([{ match: /api\.openai\.com/, status: 200, body: {} }]);
+    const adapter = new OpenAiPassthroughAdapter('openai', { fetch: fetchFn as unknown as typeof fetch });
+    await adapter.infer(makeCtx());
+    const [url] = fetchFn.mock.calls[0] as [string, RequestInit];
+    expect(url).toBe('https://api.openai.com/v1/chat/completions');
+  });
+
+  it('infer: throws for vllm when url is empty (no default)', async () => {
+    const adapter = new OpenAiPassthroughAdapter('vllm', { fetch: vi.fn() as unknown as typeof fetch });
+    await expect(adapter.infer(makeCtx())).rejects.toThrow(/no default endpoint/);
+  });
+
+  it('infer: omits Authorization when apiKey is empty', async () => {
+    const fetchFn = mockFetch([{ match: /ollama/, status: 200, body: {} }]);
+    const adapter = new OpenAiPassthroughAdapter('ollama', { fetch: fetchFn as unknown as typeof fetch });
+    await adapter.infer(makeCtx({ url: 'http://ollama:11434', apiKey: '' }));
+    const [, init] = fetchFn.mock.calls[0] as [string, RequestInit];
+    const headers = init.headers as Record<string, string>;
+    expect(headers['Authorization']).toBeUndefined();
+  });
+
+  it('stream: forwards SSE chunks and emits terminal [DONE]', async () => {
+    const fetchFn = vi.fn(async () => sseResponse([
+      'data: {"choices":[{"delta":{"content":"hi"}}]}',
+      'data: {"choices":[{"delta":{"content":"!"}}]}',
+      'data: [DONE]',
+    ]));
+    const adapter = new OpenAiPassthroughAdapter('openai', { fetch: fetchFn as unknown as typeof fetch });
+    const ctx = makeCtx({ url: 'http://example', body: { model: '', messages: [], stream: true } });
+    const chunks: { data: string; done?: boolean }[] = [];
+    for await (const c of adapter.stream(ctx)) chunks.push(c);
+    expect(chunks).toHaveLength(3);
+    expect(chunks[2]?.done).toBe(true);
+  });
+});
+
+describe('AnthropicAdapter', () => {
+  it('infer: translates system+user messages, posts to /v1/messages', async () => {
+    const fetchFn = mockFetch([{
+      match: /\/v1\/messages$/,
+      status: 200,
+      body: {
+        id: 'msg_01', model: 'claude-3-5-sonnet-20241022', role: 'assistant',
+        content: [{ type: 'text', text: 'howdy' }],
+        stop_reason: 'end_turn',
+        usage: { input_tokens: 5, output_tokens: 2 },
+      },
+    }]);
+    const adapter = new AnthropicAdapter({ fetch: fetchFn as unknown as typeof fetch });
+    const ctx = makeCtx({
+      body: {
+        model: '', messages: [
+          { role: 'system', content: 'be nice' },
+          { role: 'user', content: 'hi' },
+        ],
+      },
+      modelOverride: 'claude-3-5-sonnet-20241022',
+    });
+    const res = await adapter.infer(ctx);
+    expect(res.status).toBe(200);
+
+    const [url, init] = fetchFn.mock.calls[0] as [string, RequestInit];
+    expect(url).toBe('https://api.anthropic.com/v1/messages');
+    const headers = init.headers as Record<string, string>;
+    expect(headers['x-api-key']).toBe('test-key');
+    expect(headers['anthropic-version']).toBeDefined();
+
+    const sent = JSON.parse(init.body as string) as {
+      model: string; system: string; messages: Array<{ role: string; content: string }>; max_tokens: number;
+    };
+    expect(sent.model).toBe('claude-3-5-sonnet-20241022');
+    expect(sent.system).toBe('be nice');
+    expect(sent.messages).toEqual([{ role: 'user', content: 'hi' }]);
+    expect(sent.max_tokens).toBe(1024); // default
+
+    // Response shape: OpenAI chat.completion
+    const body = res.body as { choices: Array<{ message: { content: string }; finish_reason: string }>; usage: { total_tokens: number } };
+    expect(body.choices[0]!.message.content).toBe('howdy');
+    expect(body.choices[0]!.finish_reason).toBe('stop');
+    expect(body.usage.total_tokens).toBe(7);
+  });
+
+  it('infer: returns a synthetic error body on non-2xx', async () => {
+    const fetchFn = vi.fn(async () => new Response('boom', { status: 500 }));
+    const adapter = new AnthropicAdapter({ fetch: fetchFn as unknown as typeof fetch });
+    const res = await adapter.infer(makeCtx({ body: { model: '', messages: [{ role: 'user', content: 'x' }] } }));
+    expect(res.status).toBe(500);
+    const body = res.body as { error: { message: string } };
+    expect(body.error.message).toMatch(/HTTP 500/);
+  });
+
+  it('stream: translates anthropic event stream into OpenAI chunks', async () => {
+    const events = [
+      'event: message_start\ndata: {"type":"message_start","message":{"id":"m","content":[]}}',
+      'event: content_block_delta\ndata: {"type":"content_block_delta","delta":{"type":"text_delta","text":"he"}}',
+      'event: content_block_delta\ndata: {"type":"content_block_delta","delta":{"type":"text_delta","text":"llo"}}',
+      'event: message_delta\ndata: {"type":"message_delta","delta":{"stop_reason":"end_turn"}}',
+      'event: message_stop\ndata: {"type":"message_stop"}',
+    ];
+    const fetchFn = vi.fn(async () => sseResponse(events));
+    const adapter = new AnthropicAdapter({ fetch: fetchFn as unknown as typeof fetch });
+    const ctx = makeCtx({ body: { model: '', messages: [{ role: 'user', content: 'hi' }], stream: true } });
+
+    const chunks: { data: string; done?: boolean }[] = [];
+    for await (const c of adapter.stream(ctx)) chunks.push(c);
+
+    // Expect: role-prime, two text deltas, finish-reason, [DONE]
+    expect(chunks[chunks.length - 1]?.data).toBe('[DONE]');
+    expect(chunks[chunks.length - 1]?.done).toBe(true);
+
+    // First chunk is the role-prime (role: assistant, content: '').
+    const first = JSON.parse(chunks[0]!.data) as { choices: [{ delta: { role: string; content: string } }] };
+    expect(first.choices[0]!.delta.role).toBe('assistant');
+
+    // Next two chunks carry the text.
+    const d1 = JSON.parse(chunks[1]!.data) as { choices: [{ delta: { content: string } }] };
+    const d2 = JSON.parse(chunks[2]!.data) as { choices: [{ delta: { content: string } }] };
+    expect(d1.choices[0]!.delta.content).toBe('he');
+    expect(d2.choices[0]!.delta.content).toBe('llo');
+
+    // Finish-reason chunk.
+    const stopped = JSON.parse(chunks[3]!.data) as { choices: [{ finish_reason: string }] };
+    expect(stopped.choices[0]!.finish_reason).toBe('stop');
+  });
+});
+
+describe('LlmAdapterRegistry', () => {
+  it('returns the right adapter kind for each type', () => {
+    const reg = new LlmAdapterRegistry();
+    expect(reg.get('openai').kind).toBe('openai');
+    expect(reg.get('vllm').kind).toBe('vllm');
+    expect(reg.get('deepseek').kind).toBe('deepseek');
+    expect(reg.get('ollama').kind).toBe('ollama');
+    expect(reg.get('anthropic').kind).toBe('anthropic');
+  });
+
+  it('caches adapters between calls', () => {
+    const reg = new LlmAdapterRegistry();
+    const a = reg.get('openai');
+    const b = reg.get('openai');
+    expect(a).toBe(b);
+  });
+
+  it('rejects unsupported providers (gemini-cli is deferred)', () => {
+    const reg = new LlmAdapterRegistry();
+    expect(() => reg.get('gemini-cli')).toThrow(UnsupportedProviderError);
+    expect(() => reg.get('bogus')).toThrow(UnsupportedProviderError);
+  });
+});

src/mcpd/tests/llm-infer-route.test.ts

@@ -0,0 +1,208 @@
+import { describe, it, expect, vi, afterEach } from 'vitest';
+import Fastify from 'fastify';
+import type { FastifyInstance } from 'fastify';
+import { registerLlmInferRoutes } from '../src/routes/llm-infer.js';
+import { LlmAdapterRegistry } from '../src/services/llm/dispatcher.js';
+import { errorHandler } from '../src/middleware/error-handler.js';
+import type { LlmView } from '../src/services/llm.service.js';
+import { NotFoundError } from '../src/services/mcp-server.service.js';
+
+let app: FastifyInstance;
+
+function makeLlmView(overrides: Partial<LlmView> = {}): LlmView {
+  return {
+    id: 'llm-1',
+    name: 'claude',
+    type: 'anthropic',
+    model: 'claude-3-5-sonnet-20241022',
+    url: '',
+    tier: 'heavy',
+    description: '',
+    apiKeyRef: { name: 'anthropic-key', key: 'token' },
+    extraConfig: {},
+    version: 1,
+    createdAt: new Date(),
+    updatedAt: new Date(),
+    ...overrides,
+  };
+}
+
+afterEach(async () => {
+  if (app) await app.close();
+});
+
+function sseResponse(events: string[]): Response {
+  const body = events.join('\n\n') + '\n\n';
+  const stream = new ReadableStream<Uint8Array>({
+    start(controller) {
+      controller.enqueue(new TextEncoder().encode(body));
+      controller.close();
+    },
+  });
+  return new Response(stream, { status: 200 });
+}
+
+interface LlmServiceLike {
+  getByName: (name: string) => Promise<LlmView>;
+  resolveApiKey: (name: string) => Promise<string>;
+}
+
+async function setupApp(
+  llmService: LlmServiceLike,
+  adapters: LlmAdapterRegistry,
+  onInferenceEvent?: Parameters<typeof registerLlmInferRoutes>[1]['onInferenceEvent'],
+): Promise<FastifyInstance> {
+  app = Fastify({ logger: false });
+  app.setErrorHandler(errorHandler);
+  const deps: Parameters<typeof registerLlmInferRoutes>[1] = {
+    // eslint-disable-next-line @typescript-eslint/no-explicit-any
+    llmService: llmService as any,
+    adapters,
+  };
+  if (onInferenceEvent !== undefined) deps.onInferenceEvent = onInferenceEvent;
+  registerLlmInferRoutes(app, deps);
+  await app.ready();
+  return app;
+}
+
+describe('POST /api/v1/llms/:name/infer', () => {
+  it('returns 404 when the Llm does not exist', async () => {
+    const svc: LlmServiceLike = {
+      getByName: async () => { throw new NotFoundError('Llm not found: missing'); },
+      resolveApiKey: async () => '',
+    };
+    await setupApp(svc, new LlmAdapterRegistry());
+    const res = await app.inject({
+      method: 'POST',
+      url: '/api/v1/llms/missing/infer',
+      payload: { messages: [{ role: 'user', content: 'hi' }] },
+    });
+    expect(res.statusCode).toBe(404);
+  });
+
+  it('returns 400 when messages is missing', async () => {
+    const svc: LlmServiceLike = {
+      getByName: async () => makeLlmView({ apiKeyRef: null }),
+      resolveApiKey: async () => '',
+    };
+    await setupApp(svc, new LlmAdapterRegistry());
+    const res = await app.inject({
+      method: 'POST',
+      url: '/api/v1/llms/claude/infer',
+      payload: {},
+    });
+    expect(res.statusCode).toBe(400);
+  });
+
+  it('dispatches non-streaming to the adapter and returns its JSON', async () => {
+    const fetchFn = vi.fn(async () => new Response(JSON.stringify({
+      id: 'msg_1', model: 'claude-3-5-sonnet-20241022', role: 'assistant',
+      content: [{ type: 'text', text: 'hello' }],
+      stop_reason: 'end_turn',
+      usage: { input_tokens: 1, output_tokens: 1 },
+    }), { status: 200 }));
+    const adapters = new LlmAdapterRegistry({ fetch: fetchFn as unknown as typeof fetch });
+    const svc: LlmServiceLike = {
+      getByName: async () => makeLlmView(),
+      resolveApiKey: async () => 'sk-ant-xyz',
+    };
+    const events: unknown[] = [];
+    await setupApp(svc, adapters, (e) => events.push(e));
+
+    const res = await app.inject({
+      method: 'POST',
+      url: '/api/v1/llms/claude/infer',
+      payload: { messages: [{ role: 'user', content: 'hi' }] },
+    });
+    expect(res.statusCode).toBe(200);
+    const body = res.json<{ choices: Array<{ message: { content: string } }> }>();
+    expect(body.choices[0]!.message.content).toBe('hello');
+
+    // Audit event emitted
+    expect(events).toHaveLength(1);
+    expect((events[0] as { kind: string; llmName: string; status: number }).kind).toBe('llm_inference_call');
+    expect((events[0] as { llmName: string }).llmName).toBe('claude');
+    expect((events[0] as { streaming: boolean }).streaming).toBe(false);
+    expect((events[0] as { status: number }).status).toBe(200);
+  });
+
+  it('500s when apiKey resolution fails', async () => {
+    const adapters = new LlmAdapterRegistry();
+    const svc: LlmServiceLike = {
+      getByName: async () => makeLlmView(),
+      resolveApiKey: async () => { throw new Error('secret not found'); },
+    };
+    await setupApp(svc, adapters);
+
+    const res = await app.inject({
+      method: 'POST',
+      url: '/api/v1/llms/claude/infer',
+      payload: { messages: [{ role: 'user', content: 'hi' }] },
+    });
+    expect(res.statusCode).toBe(500);
+    expect(res.json<{ error: string }>().error).toMatch(/secret not found/);
+  });
+
+  it('skips apiKey resolution when the Llm has no apiKeyRef', async () => {
+    const fetchFn = vi.fn(async () => new Response(JSON.stringify({ id: 'x', choices: [] }), { status: 200 }));
+    const adapters = new LlmAdapterRegistry({ fetch: fetchFn as unknown as typeof fetch });
+    const resolveSpy = vi.fn();
+    const svc: LlmServiceLike = {
+      getByName: async () => makeLlmView({ type: 'ollama', url: 'http://ollama:11434', apiKeyRef: null }),
+      resolveApiKey: resolveSpy as unknown as LlmServiceLike['resolveApiKey'],
+    };
+    await setupApp(svc, adapters);
+
+    const res = await app.inject({
+      method: 'POST',
+      url: '/api/v1/llms/ollama-local/infer',
+      payload: { messages: [{ role: 'user', content: 'hi' }] },
+    });
+    expect(res.statusCode).toBe(200);
+    expect(resolveSpy).not.toHaveBeenCalled();
+  });
+
+  it('streams SSE chunks for stream: true', async () => {
+    const fetchFn = vi.fn(async () => sseResponse([
+      'event: content_block_delta\ndata: {"type":"content_block_delta","delta":{"type":"text_delta","text":"hi"}}',
+      'event: message_stop\ndata: {"type":"message_stop"}',
+    ]));
+    const adapters = new LlmAdapterRegistry({ fetch: fetchFn as unknown as typeof fetch });
+    const svc: LlmServiceLike = {
+      getByName: async () => makeLlmView(),
+      resolveApiKey: async () => 'sk-ant-xyz',
+    };
+    const events: Array<{ streaming: boolean; status: number }> = [];
+    // eslint-disable-next-line @typescript-eslint/no-explicit-any
+    await setupApp(svc, adapters, ((e: any) => events.push(e)) as any);
+
+    const res = await app.inject({
+      method: 'POST',
+      url: '/api/v1/llms/claude/infer',
+      payload: { messages: [{ role: 'user', content: 'hi' }], stream: true },
+    });
+    expect(res.statusCode).toBe(200);
+    expect(res.body).toContain('data:');
+    expect(res.body).toContain('[DONE]');
+    expect(events).toHaveLength(1);
+    expect(events[0]!.streaming).toBe(true);
+  });
+
+  it('502s on adapter errors (non-streaming)', async () => {
+    const fetchFn = vi.fn(async () => { throw new Error('upstream down'); });
+    const adapters = new LlmAdapterRegistry({ fetch: fetchFn as unknown as typeof fetch });
+    const svc: LlmServiceLike = {
+      getByName: async () => makeLlmView({ type: 'openai', url: 'http://example', apiKeyRef: null }),
+      resolveApiKey: async () => '',
+    };
+    await setupApp(svc, adapters);
+
+    const res = await app.inject({
+      method: 'POST',
+      url: '/api/v1/llms/x/infer',
+      payload: { messages: [{ role: 'user', content: 'hi' }] },
+    });
+    expect(res.statusCode).toBe(502);
+    expect(res.json<{ error: string }>().error).toMatch(/upstream down/);
+  });
+});