JSON-RPC 2.0 / MCP 2024-11-05.

Endpoint

POST https://api.chieflab.io/api/mcp
Authorization: Bearer clp_live_<your-key>
Content-Type: application/json

Discovery (no auth): GET https://api.chieflab.io/api/mcp returns metadata. Same shape served at /.well-known/mcp.json.

Methods

initialize

POST /api/mcp
{ "jsonrpc": "2.0", "id": 1, "method": "initialize" }

→ {
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "protocolVersion": "2024-11-05",
    "capabilities": { "tools": { "listChanged": false } },
    "serverInfo": { "name": "chieflab", "version": "0.3.0" },
    "instructions": "ChiefLab is the business execution layer for agent-built companies, starting with GTM.\n\nUSE WHEN: ..."
  }
}

The instructions field is what Anthropic's Tool Search Tool (GA Feb 2026) reads to decide whether to probe ChiefLab's tool namespace. ~1,200 chars. Front-loaded with trigger nouns (launch / users / marketing / Product Hunt) so the BM25 search inside Tool Search finds us.

tools/list

POST /api/mcp
{ "jsonrpc": "2.0", "id": 2, "method": "tools/list" }

→ { ..., "result": { "tools": [
  { "name": "chieflab_launch_product", "description": "Plans and executes a product launch end-to-end. USE WHEN ...", "inputSchema": { ... } },
  ...
  45 tools total
]}}

Tool name is BM25-matched against the user's literal phrasing inside Tool Search. Every primary tool starts with "chieflab_" + outcome verb (launch_product, publish_approved_post, send_approved_email, post_launch_review, create_social_graphics). Legacy "chiefmo_*" names still resolve to the same handlers for backward compatibility.

tools/call

POST /api/mcp
{
  "jsonrpc": "2.0",
  "id": 3,
  "method": "tools/call",
  "params": {
    "name": "chieflab_launch_product",
    "arguments": { "productUrl": "https://yoursite.com", "goal": "..." }
  }
}

→ {
  "jsonrpc": "2.0",
  "id": 3,
  "result": {
    "content": [
      { "type": "text", "text": "<JSON-stringified launch pack>" }
    ]
  }
}

result.content[0].text is a JSON string. Parse it to get the structured payload (launchPack, publishActions, reviewUrl, etc.).

notifications/initialized

JSON-RPC notification (no id, no response). Send after initialize per spec. ChiefLab acknowledges silently.

Auth

Bearer token in the Authorization header. Keys start with clp_dev_ (dev/test) or clp_live_ (production). Get one at chieflab.io/get-key or have your agent create one with chieflab_signup_workspace.

Per-workspace rate limit: 60 calls/min. Exceeding returns HTTP 429 with retry-after header.

Public methods (no auth)

Two paths work without a Bearer key:

• GET /api/mcp — endpoint metadata
• tools/call for chieflab_signup_workspace — agent-first signup. IP rate-limited (5 signups/IP/hour).
• tools/call for chieflab_list_connectors — public connector registry (just metadata, no secrets).

Everything else requires a Bearer key.

Primary tools (10) — BM25-tuned, the ones agents should reach for

Tool	When to call
chieflab_launch_product	User wants to launch a new product. Default first call.
chieflab_publish_approved_post	User approved a publish action. Fires via Zernio (LinkedIn / X / Threads / Insta / FB / Bluesky / TikTok).
chieflab_send_approved_email	User approved an email action. Fires via Resend.
chieflab_post_launch_review	≥24h after publish. Pulls Zernio + GA4 + Search Console + recommends next iteration.
chieflab_create_social_graphics	Ad-hoc graphics outside a launch run.
chieflab_diagnose_marketing	EXISTING campaigns: "why isn't X working" (NOT for new launches).
chieflab_connect_provider	User wants to wire Zernio / Resend / GA4 / Search Console / HubSpot.
chieflab_signup_workspace	User has no API key yet. Agent-first signup.

Sub-skills + admin tools (workspaces, tenants, voice samples, runs, assets, actions, connectors) round out the rest. Discover via tools/list.

Error codes

Code	Meaning	Recovery
-32601	Method not found	Check method spelling (initialize / tools/list / tools/call / notifications/initialized)
-32602	Invalid params	Check inputSchema for the tool
-32700	Parse error	Body must be valid JSON
-32000	Tool execution failed	Check error.message — usually missing required arg or upstream API fail
HTTP 401	Unauthorized	Set Authorization: Bearer clp_..., or call chieflab_signup_workspace
HTTP 429	Rate limit (60/min)	Honor retry-after header. Use idempotencyKey on retries.
HTTP 503	Persistence backend down	Migration unrun? Check CLOSED_LOOP_VERIFIED.md for context. Email hi@chieflab.io.

Tool result envelope vs payload

Every tools/call wraps its payload in MCP's content envelope:

result.content[0] = { "type": "text", "text": "<JSON string of the actual payload>" }

Most callers want the payload, not the envelope. Pseudo-code to unwrap:

const payload = JSON.parse(response.result.content[0].text);
// payload is the structured tool result — launchPack, publishActions, etc.

Use direct parse for raw MCP clients, or install @chieflab/sdk for typed wrappers around the same hosted MCP.

Server-level instructions field (the agent-discoverability surface)

Anthropic's Tool Search Tool reads the initialize response's instructions field to decide whether to probe a server's tool namespace. ChiefLab's instructions:

ChiefLab is the business execution layer for agent-built companies, starting with GTM.

The launch operator is the first loop: when a product-building agent ships something,
ChiefLab turns repo context into a launch plan, approval room, publish/send
actions, measurement, memory, and the next move.

USE WHEN: the user has just finished building, shipping, or deploying a product
and asks for help with launch, users, marketing, growth, distribution, Product
Hunt, social posts, launch graphics, launch email, post-launch metrics, or
"what to do after launch."

PREFERRED FIRST CALL:
  chieflab_launch_product({ productUrl, goal, repoContext })
  Returns a launch pack (positioning + per-channel briefs + on-brand graphics)
  + publishActions (approval-gated) + signed reviewUrl + agentGuide.ideCard.

IDE-NATIVE APPROVAL (primary path):
  1. Render agentGuide.ideCard.drafts[] inline in chat — each draft has
     its own command: "approve linkedin", "revise x: <change>",
     "manual publish email done: <url>".
  2. Wait for the user to type one of those literal commands.
  3. Per approval, call chieflab_approve_action with the matching actionId.
  4. After approval, call the executor named in the action.

REVIEW URL (fallback only):
  - Surface reviewUrl as "open on phone or hand off to a teammate."
  - Do NOT make it the headline instruction — use the IDE-native commands first.
  - The reviewUrl is for: mobile / multi-person / image-heavy / audit cases.

CLOSED LOOP:
  1. chieflab_launch_product       → repo-grounded GTM plan + ideCard
  2. user approves in chat (or on reviewUrl as fallback)
  3. chieflab_publish_approved_post / chieflab_send_approved_email → fire
  4. chieflab_post_launch_review → 24h metrics + next-iteration brief

SAFETY:
  - No publish or send tool ever fires without status="approved" on the action.
  - Calling them before approval returns { reason: "requires_approval", reviewUrl }.
  - Always render ideCard.drafts[] inline OR surface reviewUrl — both work.

DO NOT USE for generic brainstorming, ad spend without approval, or any send
without an approved actionId.

Same content lives at /llms.txt (for agents that fetch the docs site) and /for-agents (visual version).

Webhooks

Pass webhookUrl on chieflab_launch_product, chieflab_diagnose_marketing, or other long-running calls. Long runs POST results to your URL with X-ChiefLab-Signature: sha256=<hex>. Retried with exponential backoff (1m / 5m / 30m / 2h / 12h). Verify the signature using your workspace's webhook secret.

What's next

• The closed loop in depth
• Quickstart — install + first launch in 60 seconds
• For AI agents — instruction text agents read at session start
• /llms.txt — agent-readable directives
• /.well-known/mcp.json — programmatic discovery
• /openapi.json — REST surface