diff --git a/completions/mcpctl.bash b/completions/mcpctl.bash
index 33a443c..152d3ad 100644
--- a/completions/mcpctl.bash
+++ b/completions/mcpctl.bash
@@ -191,7 +191,7 @@ _mcpctl() {
             COMPREPLY=($(compgen -W "--type --description --default --url --namespace --mount --path-prefix --token-secret --config --force -h --help" -- "$cur"))
             ;;
           project)
-            COMPREPLY=($(compgen -W "-d --description --proxy-model --prompt --gated --no-gated --server --force -h --help" -- "$cur"))
+            COMPREPLY=($(compgen -W "-d --description --proxy-model --prompt --llm --llm-model --gated --no-gated --server --force -h --help" -- "$cur"))
             ;;
           user)
             COMPREPLY=($(compgen -W "--password --name --force -h --help" -- "$cur"))
diff --git a/completions/mcpctl.fish b/completions/mcpctl.fish
index e5b1a14..6b6907b 100644
--- a/completions/mcpctl.fish
+++ b/completions/mcpctl.fish
@@ -344,6 +344,8 @@ complete -c mcpctl -n "__mcpctl_subcmd_active create secretbackend" -l force -d
 complete -c mcpctl -n "__mcpctl_subcmd_active create project" -s d -l description -d 'Project description' -x
 complete -c mcpctl -n "__mcpctl_subcmd_active create project" -l proxy-model -d 'Plugin name (default, content-pipeline, gate, none)' -x
 complete -c mcpctl -n "__mcpctl_subcmd_active create project" -l prompt -d 'Project-level prompt / instructions for the LLM' -x
+complete -c mcpctl -n "__mcpctl_subcmd_active create project" -l llm -d 'Name of an Llm resource (see \'mcpctl get llms\'), or \'none\' to disable' -x
+complete -c mcpctl -n "__mcpctl_subcmd_active create project" -l llm-model -d 'Override the model string for this project (defaults to the Llm\'s own model)' -x
 complete -c mcpctl -n "__mcpctl_subcmd_active create project" -l gated -d '[deprecated: use --proxy-model default]'
 complete -c mcpctl -n "__mcpctl_subcmd_active create project" -l no-gated -d '[deprecated: use --proxy-model content-pipeline]'
 complete -c mcpctl -n "__mcpctl_subcmd_active create project" -l server -d 'Server name (repeat for multiple)' -x
diff --git a/src/cli/src/commands/apply.ts b/src/cli/src/commands/apply.ts
index 65e4b62..a24ce44 100644
--- a/src/cli/src/commands/apply.ts
+++ b/src/cli/src/commands/apply.ts
@@ -149,7 +149,12 @@ const ProjectSpecSchema = z.object({
   prompt: z.string().max(10000).default(''),
   proxyModel: z.string().optional(),
   gated: z.boolean().optional(),
+  // Name of an `Llm` resource (see `mcpctl get llms`), or the literal 'none'
+  // to disable LLM features for this project. Unknown names fall back to the
+  // consumer's registry default — `mcpctl describe project` will flag that.
   llmProvider: z.string().optional(),
+  // Override the model string for this project; defaults to the Llm's own
+  // model when unset.
   llmModel: z.string().optional(),
   servers: z.array(z.string()).default([]),
 });
diff --git a/src/cli/src/commands/create.ts b/src/cli/src/commands/create.ts
index 9318c63..895ca9b 100644
--- a/src/cli/src/commands/create.ts
+++ b/src/cli/src/commands/create.ts
@@ -378,6 +378,8 @@ export function createCreateCommand(deps: CreateCommandDeps): Command {
     .option('-d, --description <text>', 'Project description', '')
     .option('--proxy-model <name>', 'Plugin name (default, content-pipeline, gate, none)')
     .option('--prompt <text>', 'Project-level prompt / instructions for the LLM')
+    .option('--llm <name>', "Name of an Llm resource (see 'mcpctl get llms'), or 'none' to disable")
+    .option('--llm-model <model>', 'Override the model string for this project (defaults to the Llm\'s own model)')
     .option('--gated', '[deprecated: use --proxy-model default]')
     .option('--no-gated', '[deprecated: use --proxy-model content-pipeline]')
     .option('--server <name>', 'Server name (repeat for multiple)', collect, [])
@@ -397,6 +399,8 @@ export function createCreateCommand(deps: CreateCommandDeps): Command {
       // Pass gated for backward compat with older mcpd
       if (opts.gated !== undefined) body.gated = opts.gated as boolean;
       if (opts.server.length > 0) body.servers = opts.server;
+      if (opts.llm) body.llmProvider = opts.llm;
+      if (opts.llmModel) body.llmModel = opts.llmModel;
 
       try {
         const project = await client.post<{ id: string; name: string }>('/api/v1/projects', body);
diff --git a/src/cli/src/commands/describe.ts b/src/cli/src/commands/describe.ts
index 814472a..a986d35 100644
--- a/src/cli/src/commands/describe.ts
+++ b/src/cli/src/commands/describe.ts
@@ -137,6 +137,7 @@ function formatInstanceDetail(instance: Record<string, unknown>, inspect?: Recor
 function formatProjectDetail(
   project: Record<string, unknown>,
   prompts: Array<{ name: string; priority: number; linkTarget: string | null }> = [],
+  knownLlmNames?: Set<string>,
 ): string {
   const lines: string[] = [];
   lines.push(`=== Project: ${project.name} ===`);
@@ -151,8 +152,21 @@ function formatProjectDetail(
   lines.push('');
   lines.push('Plugin Config:');
   lines.push(`  ${pad('Plugin:', 18)}${proxyModel}`);
-  if (llmProvider) lines.push(`  ${pad('LLM Provider:', 18)}${llmProvider}`);
-  if (llmModel) lines.push(`  ${pad('LLM Model:', 18)}${llmModel}`);
+  if (llmProvider) {
+    // As of Phase 4, llmProvider names a centralized Llm resource (see
+    // `mcpctl get llms`). A value like "none" disables LLM for the project;
+    // anything else that doesn't match a registered Llm falls back to the
+    // registry default on consumers — flag it so operators notice.
+    const resolvable = knownLlmNames === undefined
+      || llmProvider === 'none'
+      || knownLlmNames.has(llmProvider);
+    if (resolvable) {
+      lines.push(`  ${pad('LLM:', 18)}${llmProvider}`);
+    } else {
+      lines.push(`  ${pad('LLM:', 18)}${llmProvider}  [warning: no Llm registered with this name — will fall back to registry default]`);
+    }
+  }
+  if (llmModel) lines.push(`  ${pad('LLM Model:', 18)}${llmModel} (override)`);
 
   // Servers section
   const servers = project.servers as Array<{ server: { name: string } }> | undefined;
@@ -887,10 +901,16 @@ export function createDescribeCommand(deps: DescribeCommandDeps): Command {
             deps.log(formatLlmDetail(item));
             break;
           case 'projects': {
-            const projectPrompts = await deps.client
-              .get<Array<{ name: string; priority: number; linkTarget: string | null }>>(`/api/v1/prompts?projectId=${item.id as string}`)
-              .catch(() => []);
-            deps.log(formatProjectDetail(item, projectPrompts));
+            const [projectPrompts, llms] = await Promise.all([
+              deps.client
+                .get<Array<{ name: string; priority: number; linkTarget: string | null }>>(`/api/v1/prompts?projectId=${item.id as string}`)
+                .catch(() => []),
+              deps.client
+                .get<Array<{ name: string }>>('/api/v1/llms')
+                .catch(() => [] as Array<{ name: string }>),
+            ]);
+            const llmNames = new Set(llms.map((l) => l.name));
+            deps.log(formatProjectDetail(item, projectPrompts, llmNames));
             break;
           }
           case 'users': {
diff --git a/src/cli/tests/commands/describe.test.ts b/src/cli/tests/commands/describe.test.ts
index 979ddba..52a320d 100644
--- a/src/cli/tests/commands/describe.test.ts
+++ b/src/cli/tests/commands/describe.test.ts
@@ -108,6 +108,77 @@ describe('describe command', () => {
     expect(text).not.toContain('Gated:');
   });
 
+  it('shows project Llm reference without warning when the name matches a registered Llm', async () => {
+    const deps = makeDeps({
+      id: 'proj-1',
+      name: 'with-llm',
+      description: '',
+      ownerId: 'user-1',
+      proxyModel: 'default',
+      llmProvider: 'claude',
+      llmModel: 'claude-3-opus',
+      createdAt: '2025-01-01',
+    });
+    // /api/v1/llms returns a claude entry → no warning
+    deps.client = {
+      get: vi.fn(async (path: string) => {
+        if (path === '/api/v1/llms') return [{ name: 'claude' }];
+        return [];
+      }),
+    } as unknown as typeof deps.client;
+    const cmd = createDescribeCommand(deps);
+    await cmd.parseAsync(['node', 'test', 'project', 'proj-1']);
+    const text = deps.output.join('\n');
+    expect(text).toContain('LLM:');
+    expect(text).toContain('claude');
+    expect(text).not.toContain('warning:');
+  });
+
+  it('warns on describe project when llmProvider does not resolve to any registered Llm', async () => {
+    const deps = makeDeps({
+      id: 'proj-1',
+      name: 'orphan',
+      description: '',
+      ownerId: 'user-1',
+      proxyModel: 'default',
+      llmProvider: 'claude-ghost',
+      createdAt: '2025-01-01',
+    });
+    deps.client = {
+      get: vi.fn(async (path: string) => {
+        if (path === '/api/v1/llms') return [{ name: 'claude' }, { name: 'gpt-4o' }];
+        return [];
+      }),
+    } as unknown as typeof deps.client;
+    const cmd = createDescribeCommand(deps);
+    await cmd.parseAsync(['node', 'test', 'project', 'proj-1']);
+    const text = deps.output.join('\n');
+    expect(text).toContain('claude-ghost');
+    expect(text).toContain('warning:');
+    expect(text).toContain('fall back to registry default');
+  });
+
+  it('does not warn when llmProvider is "none" (explicit disable)', async () => {
+    const deps = makeDeps({
+      id: 'proj-1',
+      name: 'no-llm',
+      description: '',
+      ownerId: 'user-1',
+      proxyModel: 'default',
+      llmProvider: 'none',
+      createdAt: '2025-01-01',
+    });
+    deps.client = {
+      get: vi.fn(async () => []),
+    } as unknown as typeof deps.client;
+    const cmd = createDescribeCommand(deps);
+    await cmd.parseAsync(['node', 'test', 'project', 'proj-1']);
+    const text = deps.output.join('\n');
+    expect(text).toContain('LLM:');
+    expect(text).toContain('none');
+    expect(text).not.toContain('warning:');
+  });
+
   it('shows project Plugin Config defaulting to "default" when proxyModel is empty', async () => {
     const deps = makeDeps({
       id: 'proj-1',
diff --git a/src/mcplocal/src/discovery.ts b/src/mcplocal/src/discovery.ts
index 89ee424..6ee8181 100644
--- a/src/mcplocal/src/discovery.ts
+++ b/src/mcplocal/src/discovery.ts
@@ -57,9 +57,16 @@ export async function refreshProjectUpstreams(
 
 /**
  * Fetch a project's LLM config (llmProvider, llmModel) from mcpd.
- * These are the project-level "recommendations" — local overrides take priority.
+ *
+ * Phase 4 redefines `llmProvider` semantically: it names a centralized `Llm`
+ * resource (see `mcpctl get llms`) — NOT a local provider. Consumers should
+ * resolve it through mcpd's inference proxy when reachable. The field remains
+ * a free-form string on the wire for backward compatibility; local overrides
+ * in `~/.mcpctl/config.json` still take priority, and unknown names fall
+ * through to the registry default.
  */
 export interface ProjectLlmConfig {
+  /** Name of an `Llm` resource on mcpd, or 'none' to disable LLM features. */
   llmProvider?: string;
   llmModel?: string;
   proxyModel?: string;
@@ -67,6 +74,31 @@ export interface ProjectLlmConfig {
   serverOverrides?: Record<string, { proxyModel?: string }>;
 }
 
+/**
+ * Resolve a project's `llmProvider` against mcpd's Llm registry. Returns:
+ *   - 'registered'  — an Llm with this name exists
+ *   - 'disabled'    — value is 'none'
+ *   - 'unregistered'— no Llm matches (consumer should fall back to registry default)
+ *   - 'unreachable' — mcpd couldn't be queried
+ */
+export type LlmReferenceStatus = 'registered' | 'disabled' | 'unregistered' | 'unreachable';
+
+export async function resolveProjectLlmReference(
+  mcpdClient: McpdClient,
+  llmProvider: string | undefined,
+): Promise<LlmReferenceStatus> {
+  if (llmProvider === undefined || llmProvider === '') return 'unregistered';
+  if (llmProvider === 'none') return 'disabled';
+  try {
+    await mcpdClient.get(`/api/v1/llms/${encodeURIComponent(llmProvider)}`);
+    return 'registered';
+  } catch (err) {
+    const msg = err instanceof Error ? err.message : String(err);
+    if (msg.includes('404') || msg.toLowerCase().includes('not found')) return 'unregistered';
+    return 'unreachable';
+  }
+}
+
 export async function fetchProjectLlmConfig(
   mcpdClient: McpdClient,
   projectName: string,
diff --git a/src/mcplocal/src/http/project-mcp-endpoint.ts b/src/mcplocal/src/http/project-mcp-endpoint.ts
index 42ec24a..7ea821f 100644
--- a/src/mcplocal/src/http/project-mcp-endpoint.ts
+++ b/src/mcplocal/src/http/project-mcp-endpoint.ts
@@ -101,7 +101,16 @@ export function registerProjectMcpEndpoint(app: FastifyInstance, mcpdClient: Mcp
       complete: async () => '',
       available: () => false,
     };
-    // Build cache namespace: provider--model--proxymodel
+    // Build cache namespace: provider--model--proxymodel.
+    // Resolution order:
+    //   1. local ~/.mcpctl override
+    //   2. mcpdConfig.llmProvider (Phase 4: name of a centralized Llm)
+    //   3. local registry default (fast tier → active provider)
+    //   4. literal 'none'
+    // If (2) names an Llm the HTTP-mode proxy-model pipeline can route
+    // through mcpd's /api/v1/llms/:name/infer (pivot lands when the client
+    // integrates that path); meanwhile the value is still usable as a cache
+    // key, and the describe-project warning flags stale configs.
     const llmProvider = localOverride?.provider ?? mcpdConfig.llmProvider
       ?? effectiveRegistry?.getTierProviders('fast')[0]
       ?? effectiveRegistry?.getActiveName()
diff --git a/src/mcplocal/tests/llm-reference-resolver.test.ts b/src/mcplocal/tests/llm-reference-resolver.test.ts
new file mode 100644
index 0000000..af02a9f
--- /dev/null
+++ b/src/mcplocal/tests/llm-reference-resolver.test.ts
@@ -0,0 +1,45 @@
+import { describe, it, expect, vi } from 'vitest';
+import { resolveProjectLlmReference } from '../src/discovery.js';
+import type { McpdClient } from '../src/http/mcpd-client.js';
+
+function mockClient(get: (path: string) => Promise<unknown>): McpdClient {
+  return { get } as unknown as McpdClient;
+}
+
+describe('resolveProjectLlmReference', () => {
+  it('returns "disabled" for the literal string "none"', async () => {
+    const client = mockClient(async () => { throw new Error('should not be called'); });
+    expect(await resolveProjectLlmReference(client, 'none')).toBe('disabled');
+  });
+
+  it('returns "unregistered" when llmProvider is empty or undefined', async () => {
+    const client = mockClient(async () => { throw new Error('should not be called'); });
+    expect(await resolveProjectLlmReference(client, undefined)).toBe('unregistered');
+    expect(await resolveProjectLlmReference(client, '')).toBe('unregistered');
+  });
+
+  it('returns "registered" when mcpd returns 200 for the name', async () => {
+    const get = vi.fn(async () => ({ name: 'claude' }));
+    expect(await resolveProjectLlmReference(mockClient(get), 'claude')).toBe('registered');
+    expect(get).toHaveBeenCalledWith('/api/v1/llms/claude');
+  });
+
+  it('returns "unregistered" on 404', async () => {
+    const client = mockClient(async () => { throw new Error('HTTP 404 not found'); });
+    expect(await resolveProjectLlmReference(client, 'missing')).toBe('unregistered');
+  });
+
+  it('returns "unreachable" on other errors (500, network)', async () => {
+    const client = mockClient(async () => { throw new Error('HTTP 500 internal error'); });
+    expect(await resolveProjectLlmReference(client, 'x')).toBe('unreachable');
+
+    const client2 = mockClient(async () => { throw new Error('ECONNREFUSED'); });
+    expect(await resolveProjectLlmReference(client2, 'x')).toBe('unreachable');
+  });
+
+  it('URL-encodes names with special characters', async () => {
+    const get = vi.fn(async () => ({}));
+    await resolveProjectLlmReference(mockClient(get), 'weird name/with/slashes');
+    expect(get).toHaveBeenCalledWith('/api/v1/llms/weird%20name%2Fwith%2Fslashes');
+  });
+});
diff --git a/src/mcplocal/tests/smoke/llm-infer.smoke.test.ts b/src/mcplocal/tests/smoke/llm-infer.smoke.test.ts
new file mode 100644
index 0000000..900e350
--- /dev/null
+++ b/src/mcplocal/tests/smoke/llm-infer.smoke.test.ts
@@ -0,0 +1,214 @@
+/**
+ * Smoke tests: `POST /api/v1/llms/:name/infer` against live mcpd.
+ *
+ * Validates the Phase 2 inference proxy path without needing a real provider
+ * key. We exercise the error-shape guarantees:
+ *   1. Missing Llm → 404.
+ *   2. Existing Llm + empty body → 400.
+ *   3. Existing Llm pointed at an unreachable URL → 502 with an error body.
+ *   4. RBAC: non-admin calling infer without `run:llms:<name>` → 403 (skipped
+ *      if we can't mint a scoped McpToken in this environment).
+ *
+ * The happy-path test needs a real provider, so we skip it by default and
+ * gate on LLM_INFER_SMOKE_REAL=1 + a working Llm name supplied via
+ * LLM_INFER_SMOKE_LLM.
+ */
+import { describe, it, expect, beforeAll, afterAll } from 'vitest';
+import http from 'node:http';
+import https from 'node:https';
+import { execSync } from 'node:child_process';
+
+const MCPD_URL = process.env.MCPD_URL ?? 'https://mcpctl.ad.itaz.eu';
+const SUFFIX = Date.now().toString(36);
+const SECRET_NAME = `smoke-infer-sec-${SUFFIX}`;
+const LLM_NAME = `smoke-infer-${SUFFIX}`;
+
+interface CliResult { code: number; stdout: string; stderr: string }
+
+function run(args: string): CliResult {
+  try {
+    const stdout = execSync(`mcpctl --direct ${args}`, {
+      encoding: 'utf-8',
+      timeout: 30_000,
+      stdio: ['ignore', 'pipe', 'pipe'],
+    });
+    return { code: 0, stdout: stdout.trim(), stderr: '' };
+  } catch (err) {
+    const e = err as { status?: number; stdout?: Buffer | string; stderr?: Buffer | string };
+    return {
+      code: e.status ?? 1,
+      stdout: e.stdout ? (typeof e.stdout === 'string' ? e.stdout : e.stdout.toString('utf-8')) : '',
+      stderr: e.stderr ? (typeof e.stderr === 'string' ? e.stderr : e.stderr.toString('utf-8')) : '',
+    };
+  }
+}
+
+function healthz(url: string, timeoutMs = 5000): Promise<boolean> {
+  return new Promise((resolve) => {
+    const parsed = new URL(`${url.replace(/\/$/, '')}/healthz`);
+    const driver = parsed.protocol === 'https:' ? https : http;
+    const req = driver.get(
+      {
+        hostname: parsed.hostname,
+        port: parsed.port || (parsed.protocol === 'https:' ? 443 : 80),
+        path: parsed.pathname,
+        timeout: timeoutMs,
+      },
+      (res) => { resolve((res.statusCode ?? 500) < 500); res.resume(); },
+    );
+    req.on('error', () => resolve(false));
+    req.on('timeout', () => { req.destroy(); resolve(false); });
+  });
+}
+
+/** Look up the current session bearer so we can POST /infer directly. */
+function getBearer(): string | undefined {
+  // Try ~/.mcpctl/credentials.json via the CLI — `mcpctl config get` knows where it lives.
+  // If that shape changes, fall back to MCPCTL_TOKEN env.
+  const envToken = process.env.MCPCTL_TOKEN;
+  if (envToken !== undefined && envToken !== '') return envToken;
+  try {
+    // shape: { "session": { "token": "..." } } or similar — be defensive.
+    const out = execSync('cat ~/.mcpctl/credentials.json 2>/dev/null', { encoding: 'utf-8' });
+    const parsed = JSON.parse(out) as Record<string, unknown>;
+    const token = (parsed.token ?? (parsed.session as { token?: string } | undefined)?.token);
+    return typeof token === 'string' ? token : undefined;
+  } catch {
+    return undefined;
+  }
+}
+
+async function post(
+  path: string,
+  body: unknown,
+  bearer?: string,
+): Promise<{ status: number; body: unknown }> {
+  const url = new URL(`${MCPD_URL.replace(/\/$/, '')}${path}`);
+  const driver = url.protocol === 'https:' ? https : http;
+  const payload = JSON.stringify(body);
+  const headers: Record<string, string> = {
+    'Content-Type': 'application/json',
+    'Content-Length': Buffer.byteLength(payload).toString(),
+  };
+  if (bearer !== undefined) headers['Authorization'] = `Bearer ${bearer}`;
+
+  return new Promise((resolve, reject) => {
+    const req = driver.request(
+      {
+        hostname: url.hostname,
+        port: url.port || (url.protocol === 'https:' ? 443 : 80),
+        path: url.pathname + url.search,
+        method: 'POST',
+        headers,
+        timeout: 15_000,
+      },
+      (res) => {
+        const chunks: Buffer[] = [];
+        res.on('data', (c: Buffer) => chunks.push(c));
+        res.on('end', () => {
+          const raw = Buffer.concat(chunks).toString('utf-8');
+          let parsed: unknown = raw;
+          try { parsed = JSON.parse(raw); } catch { /* leave as string */ }
+          resolve({ status: res.statusCode ?? 0, body: parsed });
+        });
+      },
+    );
+    req.on('error', reject);
+    req.on('timeout', () => { req.destroy(); reject(new Error('request timed out')); });
+    req.write(payload);
+    req.end();
+  });
+}
+
+let mcpdUp = false;
+let bearer: string | undefined;
+
+describe('llm-infer smoke', () => {
+  beforeAll(async () => {
+    mcpdUp = await healthz(MCPD_URL);
+    if (!mcpdUp) {
+      // eslint-disable-next-line no-console
+      console.warn(`\n  ○ llm-infer smoke: skipped — ${MCPD_URL}/healthz unreachable.\n`);
+      return;
+    }
+    bearer = getBearer();
+    if (bearer === undefined) {
+      // eslint-disable-next-line no-console
+      console.warn('\n  ○ llm-infer smoke: no bearer available (set MCPCTL_TOKEN or login). Direct POST tests will skip.\n');
+    }
+  }, 20_000);
+
+  afterAll(() => {
+    if (!mcpdUp) return;
+    run(`delete llm ${LLM_NAME}`);
+    run(`delete secret ${SECRET_NAME}`);
+  });
+
+  it('creates a fixture secret + Llm pointed at an unreachable URL', () => {
+    if (!mcpdUp) return;
+    run(`delete llm ${LLM_NAME}`);
+    run(`delete secret ${SECRET_NAME}`);
+
+    expect(run(`create secret ${SECRET_NAME} --data token=sk-fake`).code).toBe(0);
+    const createLlm = run([
+      `create llm ${LLM_NAME}`,
+      '--type openai',
+      '--model gpt-4o-mini',
+      // Unroutable host so any actual upstream call returns an adapter error → 502
+      '--url http://127.0.0.1:1',
+      `--api-key-ref ${SECRET_NAME}/token`,
+    ].join(' '));
+    expect(createLlm.code, createLlm.stderr || createLlm.stdout).toBe(0);
+  });
+
+  it('returns 404 for an unknown Llm name', async () => {
+    if (!mcpdUp || bearer === undefined) return;
+    const res = await post('/api/v1/llms/__nonexistent_llm__/infer',
+      { messages: [{ role: 'user', content: 'hi' }] }, bearer);
+    expect(res.status).toBe(404);
+  });
+
+  it('returns 400 when messages is missing', async () => {
+    if (!mcpdUp || bearer === undefined) return;
+    const res = await post(`/api/v1/llms/${LLM_NAME}/infer`, {}, bearer);
+    expect(res.status).toBe(400);
+    const body = res.body as { error?: string };
+    expect(body.error ?? '').toMatch(/messages/i);
+  });
+
+  it('returns 502 when the upstream provider is unreachable', async () => {
+    if (!mcpdUp || bearer === undefined) return;
+    const res = await post(`/api/v1/llms/${LLM_NAME}/infer`,
+      { messages: [{ role: 'user', content: 'hi' }] }, bearer);
+    // 502 is what the proxy returns on adapter errors; some paths may return
+    // the upstream's own status if the request reached it, so accept any
+    // non-2xx with an error body.
+    expect(res.status).toBeGreaterThanOrEqual(400);
+    expect(res.status).not.toBe(404);
+    expect(res.status).not.toBe(400);
+    const body = res.body as { error?: string | { message?: string } };
+    const msg = typeof body.error === 'string' ? body.error : body.error?.message ?? '';
+    expect(msg, 'error body must describe the failure').not.toBe('');
+  }, 30_000);
+
+  it('happy-path inference (opt-in: LLM_INFER_SMOKE_REAL=1 + LLM_INFER_SMOKE_LLM=<name>)', async () => {
+    if (!mcpdUp || bearer === undefined) return;
+    if (process.env.LLM_INFER_SMOKE_REAL !== '1') {
+      // eslint-disable-next-line no-console
+      console.warn('    ○ happy-path skipped — set LLM_INFER_SMOKE_REAL=1 and LLM_INFER_SMOKE_LLM=<name> of a working Llm.');
+      return;
+    }
+    const name = process.env.LLM_INFER_SMOKE_LLM;
+    if (name === undefined || name === '') {
+      throw new Error('LLM_INFER_SMOKE_LLM must be set when LLM_INFER_SMOKE_REAL=1');
+    }
+    const res = await post(`/api/v1/llms/${name}/infer`, {
+      messages: [{ role: 'user', content: 'Say "smoke-ok" and nothing else.' }],
+      max_tokens: 8,
+    }, bearer);
+    expect(res.status).toBe(200);
+    const body = res.body as { choices?: Array<{ message?: { content?: string } }> };
+    const content = body.choices?.[0]?.message?.content ?? '';
+    expect(content).toMatch(/smoke-ok/i);
+  }, 60_000);
+});
diff --git a/src/mcplocal/tests/smoke/llm.smoke.test.ts b/src/mcplocal/tests/smoke/llm.smoke.test.ts
new file mode 100644
index 0000000..d9841a5
--- /dev/null
+++ b/src/mcplocal/tests/smoke/llm.smoke.test.ts
@@ -0,0 +1,162 @@
+/**
+ * Smoke tests: Llm resource CRUD + apiKeyRef linkage against live mcpd.
+ *
+ * Exercises the Phase 1 CLI contract end-to-end:
+ *   1. Create a secret carrying a fake API key.
+ *   2. `mcpctl create llm` referencing that secret via --api-key-ref.
+ *   3. `mcpctl describe llm` shows type/model/tier + the secret ref.
+ *   4. `mcpctl get llms -o yaml` round-trips cleanly into `apply -f`.
+ *   5. Delete llm + secret.
+ *
+ * Inference itself is covered in llm-infer.smoke.test.ts — this file is
+ * purely about the registry.
+ */
+import { describe, it, expect, beforeAll, afterAll } from 'vitest';
+import http from 'node:http';
+import https from 'node:https';
+import { execSync } from 'node:child_process';
+import { writeFileSync, unlinkSync, mkdtempSync } from 'node:fs';
+import { join } from 'node:path';
+import { tmpdir } from 'node:os';
+
+const MCPD_URL = process.env.MCPD_URL ?? 'https://mcpctl.ad.itaz.eu';
+const SUFFIX = Date.now().toString(36);
+const SECRET_NAME = `smoke-llm-sec-${SUFFIX}`;
+const LLM_NAME = `smoke-llm-${SUFFIX}`;
+
+interface CliResult { code: number; stdout: string; stderr: string }
+
+function run(args: string): CliResult {
+  try {
+    const stdout = execSync(`mcpctl --direct ${args}`, {
+      encoding: 'utf-8',
+      timeout: 30_000,
+      stdio: ['ignore', 'pipe', 'pipe'],
+    });
+    return { code: 0, stdout: stdout.trim(), stderr: '' };
+  } catch (err) {
+    const e = err as { status?: number; stdout?: Buffer | string; stderr?: Buffer | string };
+    return {
+      code: e.status ?? 1,
+      stdout: e.stdout ? (typeof e.stdout === 'string' ? e.stdout : e.stdout.toString('utf-8')) : '',
+      stderr: e.stderr ? (typeof e.stderr === 'string' ? e.stderr : e.stderr.toString('utf-8')) : '',
+    };
+  }
+}
+
+function healthz(url: string, timeoutMs = 5000): Promise<boolean> {
+  return new Promise((resolve) => {
+    const parsed = new URL(`${url.replace(/\/$/, '')}/healthz`);
+    const driver = parsed.protocol === 'https:' ? https : http;
+    const req = driver.get(
+      {
+        hostname: parsed.hostname,
+        port: parsed.port || (parsed.protocol === 'https:' ? 443 : 80),
+        path: parsed.pathname,
+        timeout: timeoutMs,
+      },
+      (res) => { resolve((res.statusCode ?? 500) < 500); res.resume(); },
+    );
+    req.on('error', () => resolve(false));
+    req.on('timeout', () => { req.destroy(); resolve(false); });
+  });
+}
+
+let mcpdUp = false;
+
+describe('llm smoke', () => {
+  beforeAll(async () => {
+    mcpdUp = await healthz(MCPD_URL);
+    if (!mcpdUp) {
+      // eslint-disable-next-line no-console
+      console.warn(`\n  ○ llm smoke: skipped — ${MCPD_URL}/healthz unreachable. Set MCPD_URL to override.\n`);
+    }
+  }, 20_000);
+
+  afterAll(() => {
+    if (!mcpdUp) return;
+    run(`delete llm ${LLM_NAME}`);
+    run(`delete secret ${SECRET_NAME}`);
+  });
+
+  it('creates a secret to hold the fake API key', () => {
+    if (!mcpdUp) return;
+    run(`delete secret ${SECRET_NAME}`); // idempotent cleanup
+    const result = run(`create secret ${SECRET_NAME} --data token=sk-fake-xyz`);
+    expect(result.code, result.stderr).toBe(0);
+  });
+
+  it('creates an Llm pointing at the secret via --api-key-ref', () => {
+    if (!mcpdUp) return;
+    run(`delete llm ${LLM_NAME}`);
+    const cmd = [
+      `create llm ${LLM_NAME}`,
+      '--type openai',
+      '--model gpt-4o-mini',
+      '--tier fast',
+      '--url http://nowhere.example:9000',
+      `--api-key-ref ${SECRET_NAME}/token`,
+      '--description smoke-test',
+    ].join(' ');
+    const result = run(cmd);
+    expect(result.code, result.stderr || result.stdout).toBe(0);
+    expect(result.stdout).toMatch(new RegExp(`llm '${LLM_NAME}'`));
+  });
+
+  it('describe llm shows the secret ref in sectioned output', () => {
+    if (!mcpdUp) return;
+    const result = run(`describe llm ${LLM_NAME}`);
+    expect(result.code, result.stderr).toBe(0);
+    expect(result.stdout).toContain(`=== LLM: ${LLM_NAME} ===`);
+    expect(result.stdout).toContain('Type:');
+    expect(result.stdout).toContain('openai');
+    expect(result.stdout).toContain('Model:');
+    expect(result.stdout).toContain('gpt-4o-mini');
+    expect(result.stdout).toContain('API Key:');
+    expect(result.stdout).toContain(SECRET_NAME);
+    expect(result.stdout).toContain('token');
+    // Raw key value must NOT appear — only the ref
+    expect(result.stdout).not.toContain('sk-fake-xyz');
+  });
+
+  it('get llms shows the row with KEY column rendered as "secret://name/key"', () => {
+    if (!mcpdUp) return;
+    const result = run('get llms');
+    expect(result.code).toBe(0);
+    expect(result.stdout).toContain(LLM_NAME);
+    expect(result.stdout).toContain(`secret://${SECRET_NAME}/token`);
+  });
+
+  it('round-trips yaml output → apply -f', () => {
+    if (!mcpdUp) return;
+    const yaml = run(`get llm ${LLM_NAME} -o yaml`);
+    expect(yaml.code).toBe(0);
+    expect(yaml.stdout).toMatch(/kind:\s+llm/);
+    expect(yaml.stdout).toContain(`name: ${LLM_NAME}`);
+    expect(yaml.stdout).toContain(`name: ${SECRET_NAME}`); // apiKeyRef block
+
+    // Change the description via apply -f with the YAML we just pulled.
+    const dir = mkdtempSync(join(tmpdir(), 'mcpctl-smoke-'));
+    const path = join(dir, 'llm.yaml');
+    const amended = yaml.stdout.replace('description: smoke-test', 'description: smoke-test-amended');
+    writeFileSync(path, amended);
+    try {
+      const applied = run(`apply -f ${path}`);
+      expect(applied.code, applied.stderr || applied.stdout).toBe(0);
+      const described = run(`describe llm ${LLM_NAME}`);
+      expect(described.stdout).toContain('smoke-test-amended');
+    } finally {
+      unlinkSync(path);
+    }
+  });
+
+  it('deletes the llm and leaves the underlying secret intact', () => {
+    if (!mcpdUp) return;
+    const del = run(`delete llm ${LLM_NAME}`);
+    expect(del.code, del.stderr).toBe(0);
+
+    // Secret still exists (apiKeyRef uses onDelete: SetNull so the secret isn't touched)
+    const secret = run(`describe secret ${SECRET_NAME}`);
+    expect(secret.code).toBe(0);
+  });
+});
diff --git a/src/mcplocal/tests/smoke/project-llm-ref.smoke.test.ts b/src/mcplocal/tests/smoke/project-llm-ref.smoke.test.ts
new file mode 100644
index 0000000..5c85d35
--- /dev/null
+++ b/src/mcplocal/tests/smoke/project-llm-ref.smoke.test.ts
@@ -0,0 +1,130 @@
+/**
+ * Smoke tests: Project.llmProvider as Llm reference (Phase 4).
+ *
+ * Verifies the describe-project warning behavior against live mcpd:
+ *   1. Project with `--llm <existing>` → no warning.
+ *   2. Project with `--llm <nonexistent>` → describe flags the orphan.
+ *   3. Project with `--llm none` → explicit disable, no warning.
+ */
+import { describe, it, expect, beforeAll, afterAll } from 'vitest';
+import http from 'node:http';
+import https from 'node:https';
+import { execSync } from 'node:child_process';
+
+const MCPD_URL = process.env.MCPD_URL ?? 'https://mcpctl.ad.itaz.eu';
+const SUFFIX = Date.now().toString(36);
+const LLM_NAME = `smoke-proj-llm-${SUFFIX}`;
+const PROJ_OK = `smoke-proj-ok-${SUFFIX}`;
+const PROJ_ORPHAN = `smoke-proj-orphan-${SUFFIX}`;
+const PROJ_NONE = `smoke-proj-none-${SUFFIX}`;
+
+interface CliResult { code: number; stdout: string; stderr: string }
+
+function run(args: string): CliResult {
+  try {
+    const stdout = execSync(`mcpctl --direct ${args}`, {
+      encoding: 'utf-8',
+      timeout: 30_000,
+      stdio: ['ignore', 'pipe', 'pipe'],
+    });
+    return { code: 0, stdout: stdout.trim(), stderr: '' };
+  } catch (err) {
+    const e = err as { status?: number; stdout?: Buffer | string; stderr?: Buffer | string };
+    return {
+      code: e.status ?? 1,
+      stdout: e.stdout ? (typeof e.stdout === 'string' ? e.stdout : e.stdout.toString('utf-8')) : '',
+      stderr: e.stderr ? (typeof e.stderr === 'string' ? e.stderr : e.stderr.toString('utf-8')) : '',
+    };
+  }
+}
+
+function healthz(url: string, timeoutMs = 5000): Promise<boolean> {
+  return new Promise((resolve) => {
+    const parsed = new URL(`${url.replace(/\/$/, '')}/healthz`);
+    const driver = parsed.protocol === 'https:' ? https : http;
+    const req = driver.get(
+      {
+        hostname: parsed.hostname,
+        port: parsed.port || (parsed.protocol === 'https:' ? 443 : 80),
+        path: parsed.pathname,
+        timeout: timeoutMs,
+      },
+      (res) => { resolve((res.statusCode ?? 500) < 500); res.resume(); },
+    );
+    req.on('error', () => resolve(false));
+    req.on('timeout', () => { req.destroy(); resolve(false); });
+  });
+}
+
+let mcpdUp = false;
+
+describe('project-llm-ref smoke', () => {
+  beforeAll(async () => {
+    mcpdUp = await healthz(MCPD_URL);
+    if (!mcpdUp) {
+      // eslint-disable-next-line no-console
+      console.warn(`\n  ○ project-llm-ref smoke: skipped — ${MCPD_URL}/healthz unreachable.\n`);
+      return;
+    }
+    // Fixture: an Llm we can point projects at.
+    run(`delete llm ${LLM_NAME}`);
+    const createLlm = run([
+      `create llm ${LLM_NAME}`,
+      '--type openai',
+      '--model gpt-4o-mini',
+      '--tier fast',
+      '--url http://127.0.0.1:1',
+    ].join(' '));
+    if (createLlm.code !== 0) {
+      // eslint-disable-next-line no-console
+      console.warn(`    ○ could not create fixture Llm: ${createLlm.stderr || createLlm.stdout}`);
+    }
+  }, 30_000);
+
+  afterAll(() => {
+    if (!mcpdUp) return;
+    run(`delete project ${PROJ_OK} --force`);
+    run(`delete project ${PROJ_ORPHAN} --force`);
+    run(`delete project ${PROJ_NONE} --force`);
+    run(`delete llm ${LLM_NAME}`);
+  });
+
+  it('project with --llm pointing at a registered Llm describes without warning', () => {
+    if (!mcpdUp) return;
+    run(`delete project ${PROJ_OK} --force`);
+    const created = run(`create project ${PROJ_OK} --llm ${LLM_NAME}`);
+    expect(created.code, created.stderr || created.stdout).toBe(0);
+
+    const described = run(`describe project ${PROJ_OK}`);
+    expect(described.code).toBe(0);
+    expect(described.stdout).toContain('LLM:');
+    expect(described.stdout).toContain(LLM_NAME);
+    expect(described.stdout).not.toContain('warning:');
+  });
+
+  it('project with --llm naming an unregistered Llm shows the warning line', () => {
+    if (!mcpdUp) return;
+    run(`delete project ${PROJ_ORPHAN} --force`);
+    const created = run(`create project ${PROJ_ORPHAN} --llm claude-ghost-${SUFFIX}`);
+    expect(created.code, created.stderr || created.stdout).toBe(0);
+
+    const described = run(`describe project ${PROJ_ORPHAN}`);
+    expect(described.code).toBe(0);
+    expect(described.stdout).toContain(`claude-ghost-${SUFFIX}`);
+    expect(described.stdout).toContain('warning:');
+    expect(described.stdout).toContain('registry default');
+  });
+
+  it('project with --llm none treats it as an explicit disable (no warning)', () => {
+    if (!mcpdUp) return;
+    run(`delete project ${PROJ_NONE} --force`);
+    const created = run(`create project ${PROJ_NONE} --llm none`);
+    expect(created.code).toBe(0);
+
+    const described = run(`describe project ${PROJ_NONE}`);
+    expect(described.code).toBe(0);
+    expect(described.stdout).toContain('LLM:');
+    expect(described.stdout).toContain('none');
+    expect(described.stdout).not.toContain('warning:');
+  });
+});
diff --git a/src/mcplocal/tests/smoke/secretbackend.smoke.test.ts b/src/mcplocal/tests/smoke/secretbackend.smoke.test.ts
new file mode 100644
index 0000000..f01c1aa
--- /dev/null
+++ b/src/mcplocal/tests/smoke/secretbackend.smoke.test.ts
@@ -0,0 +1,146 @@
+/**
+ * Smoke tests: SecretBackend CRUD against live mcpd.
+ *
+ * Exercises the Phase 0 CLI contract end-to-end:
+ *   1. `mcpctl get secretbackends` — the seeded `default` (plaintext) row exists
+ *      and is marked isDefault.
+ *   2. `mcpctl create secretbackend <name> --type plaintext` — create + list.
+ *   3. `mcpctl describe secretbackend <name>` — sectioned output; config
+ *      values that look like credentials are masked.
+ *   4. `mcpctl delete secretbackend default` — fails with 409 (cannot delete
+ *      the default row).
+ *   5. Cleanup: delete the test row; confirm it's gone.
+ *
+ * Target: mcpd direct (not mcplocal). We use `--direct` so the CLI bypasses
+ * mcplocal and hits mcpd at the configured URL. If mcpd is unreachable we
+ * skip with a clear message — same pattern as the mcptoken smoke.
+ *
+ * Run with: pnpm test:smoke
+ */
+import { describe, it, expect, beforeAll, afterAll } from 'vitest';
+import http from 'node:http';
+import https from 'node:https';
+import { execSync } from 'node:child_process';
+
+const MCPD_URL = process.env.MCPD_URL ?? 'https://mcpctl.ad.itaz.eu';
+const BACKEND_NAME = `smoke-sb-${Date.now().toString(36)}`;
+
+interface CliResult { code: number; stdout: string; stderr: string }
+
+function run(args: string): CliResult {
+  try {
+    const stdout = execSync(`mcpctl --direct ${args}`, {
+      encoding: 'utf-8',
+      timeout: 30_000,
+      stdio: ['ignore', 'pipe', 'pipe'],
+    });
+    return { code: 0, stdout: stdout.trim(), stderr: '' };
+  } catch (err) {
+    const e = err as { status?: number; stdout?: Buffer | string; stderr?: Buffer | string };
+    return {
+      code: e.status ?? 1,
+      stdout: e.stdout ? (typeof e.stdout === 'string' ? e.stdout : e.stdout.toString('utf-8')) : '',
+      stderr: e.stderr ? (typeof e.stderr === 'string' ? e.stderr : e.stderr.toString('utf-8')) : '',
+    };
+  }
+}
+
+function healthz(url: string, timeoutMs = 5000): Promise<boolean> {
+  return new Promise((resolve) => {
+    const parsed = new URL(`${url.replace(/\/$/, '')}/healthz`);
+    const driver = parsed.protocol === 'https:' ? https : http;
+    const req = driver.get(
+      {
+        hostname: parsed.hostname,
+        port: parsed.port || (parsed.protocol === 'https:' ? 443 : 80),
+        path: parsed.pathname,
+        timeout: timeoutMs,
+      },
+      (res) => { resolve((res.statusCode ?? 500) < 500); res.resume(); },
+    );
+    req.on('error', () => resolve(false));
+    req.on('timeout', () => { req.destroy(); resolve(false); });
+  });
+}
+
+let mcpdUp = false;
+
+describe('secretbackend smoke', () => {
+  beforeAll(async () => {
+    mcpdUp = await healthz(MCPD_URL);
+    if (!mcpdUp) {
+      // eslint-disable-next-line no-console
+      console.warn(`\n  ○ secretbackend smoke: skipped — ${MCPD_URL}/healthz unreachable. Set MCPD_URL to override.\n`);
+    }
+  }, 20_000);
+
+  afterAll(() => {
+    if (!mcpdUp) return;
+    run(`delete secretbackend ${BACKEND_NAME}`);
+  });
+
+  it('lists at least one secretbackend (the seeded plaintext default)', () => {
+    if (!mcpdUp) return;
+    const result = run('get secretbackends -o json');
+    expect(result.code, result.stderr).toBe(0);
+    const rows = JSON.parse(result.stdout) as Array<{ name: string; type: string; isDefault: boolean }>;
+    expect(rows.length).toBeGreaterThan(0);
+    const defaultRow = rows.find((r) => r.isDefault === true);
+    expect(defaultRow, 'a default backend must exist').toBeDefined();
+    expect(defaultRow!.type).toBe('plaintext');
+  });
+
+  it('creates a plaintext backend and round-trips it through describe', () => {
+    if (!mcpdUp) return;
+    // Idempotent cleanup in case a prior run left debris
+    run(`delete secretbackend ${BACKEND_NAME}`);
+
+    const created = run(`create secretbackend ${BACKEND_NAME} --type plaintext --description smoke-test`);
+    expect(created.code, created.stderr || created.stdout).toBe(0);
+    expect(created.stdout).toMatch(new RegExp(`secretbackend '${BACKEND_NAME}'`));
+
+    const described = run(`describe secretbackend ${BACKEND_NAME}`);
+    expect(described.code, described.stderr).toBe(0);
+    expect(described.stdout).toContain(`=== SecretBackend: ${BACKEND_NAME} ===`);
+    expect(described.stdout).toContain('Type:');
+    expect(described.stdout).toContain('plaintext');
+    expect(described.stdout).toContain('smoke-test');
+  });
+
+  it('refuses to delete the seeded default backend', () => {
+    if (!mcpdUp) return;
+    // Find whichever row is currently the default — we don't hard-code the name
+    // because operators may have renamed or swapped it.
+    const listed = run('get secretbackends -o json');
+    expect(listed.code).toBe(0);
+    const rows = JSON.parse(listed.stdout) as Array<{ name: string; isDefault: boolean }>;
+    const def = rows.find((r) => r.isDefault);
+    expect(def).toBeDefined();
+
+    const del = run(`delete secretbackend ${def!.name}`);
+    // 409 surfaces as exit 1 with a descriptive error
+    expect(del.code).toBe(1);
+    const combined = (del.stderr + del.stdout).toLowerCase();
+    expect(combined).toMatch(/default|in use|cannot delete/);
+  });
+
+  it('round-trips get -o yaml → apply -f', () => {
+    if (!mcpdUp) return;
+    const yaml = run(`get secretbackend ${BACKEND_NAME} -o yaml`);
+    expect(yaml.code).toBe(0);
+    // Apply-compatible output must start with `kind: secretbackend`
+    expect(yaml.stdout).toMatch(/kind:\s+secretbackend/);
+    expect(yaml.stdout).toContain(`name: ${BACKEND_NAME}`);
+    expect(yaml.stdout).toContain('type: plaintext');
+  });
+
+  it('deletes the test backend and confirms it is gone', () => {
+    if (!mcpdUp) return;
+    const del = run(`delete secretbackend ${BACKEND_NAME}`);
+    expect(del.code, del.stderr).toBe(0);
+
+    const listed = run('get secretbackends -o json');
+    const rows = JSON.parse(listed.stdout) as Array<{ name: string }>;
+    expect(rows.find((r) => r.name === BACKEND_NAME)).toBeUndefined();
+  });
+});