MCP API Overview

clanker.engineer implements the Model Context Protocol (MCP), allowing AI assistants to interact with skills, artifacts, and executions through a standardized interface.

What is MCP?

The Model Context Protocol is an open standard that enables AI assistants to use external tools and access data sources. When you connect clanker to Claude Desktop, Cursor, or VS Code, your AI assistant gains access to our complete tool suite.

Endpoints

Streamable HTTP is the only MCP transport. The older SSE pair has been removed.

Authentication

All MCP endpoints require authentication via API key:

{
  "headers": {
    "x-api-key": "ck_live_xxxxxxxxxxxxx"
  }
}

A ck_ key is also accepted as Authorization: Bearer ck_…. What a key can see and call is bounded by its scopes: tools/list shows only the tools its scopes cover, and tools/call on any other tool is refused with an error naming the missing scope.

See Authentication for details on obtaining an API key.

Available Tools

Tools are organized by function:

Skills

Executions

Artifacts

Billing

Connectors

The unified connector surface — sources (code repos, files) and actions (SaaS operations) all live here. The previously-separate “Integrations” tool family was consolidated into the connector family.

Reading through a connected account (list-github-repos, list-sources, list-branches, get-source-content) also needs connector:exec; the tools that only report what is connected (list-connectors, get-connector-status, search-connector-actions, list-connector-triggers) need connector:list. A tool the credential’s scopes do not cover is hidden from tools/list. See Agent Permissions.

Workflows

Install & Resources

Agents, traces and fetching

Response Format

All tools return a standardized response:

Success Response

{
  "success": true,
  "data": { ... },
  "pagination": {
    "cursor": "...",
    "hasMore": true
  }
}

Error Response

{
  "success": false,
  "error": {
    "code": "NOT_FOUND",
    "message": "Resource not found"
  }
}

Rate Limiting

MCP requests are rate-limited to 60 requests per minute per user.

Rate limit info is returned in HTTP response headers.

Pagination

List operations support cursor-based pagination:

{
  "tool": "list-installed-skills",
  "arguments": {
    "limit": 20,
    "cursor": "previous_cursor_value"
  }
}

Response includes pagination metadata at the top level:

{
  "success": true,
  "data": [...],
  "pagination": {
    "cursor": "next_cursor_value",
    "hasMore": true
  }
}

Note: Some tools (such as list-executions) use offset-based pagination instead of cursor-based. These tools accept an offset parameter rather than a cursor.

Client Configuration

See MCP Client Setup for detailed setup instructions for:

• Claude Desktop
• Cursor
• VS Code
• Custom clients

Error Codes

Best Practices

1. Check status first - Before executing, verify no execution is running
2. Handle pagination - Always check hasMore and use cursors
3. Respect rate limits - Check x-ratelimit-remaining in response headers
4. Handle timeouts gracefully - Long-running operations may timeout