Developer MCP and skills

Connect an AI agent to Caraer developer documentation

The developer MCP lets an agent discover features, search guides, inspect API operations, and read task-focused skills. It serves public documentation and OpenAPI; it does not connect to a company or execute Caraer API calls.

Connect

Add a remote Streamable HTTP MCP server with this URL:

https://developer-mcp.caraer.com/mcp

For clients using an mcpServers configuration:

{
  "mcpServers": {
    "caraer-developer-docs": {
      "url": "https://developer-mcp.caraer.com/mcp"
    }
  }
}

No company credentials are required. Ask the agent: “Using Caraer developer docs, find the endpoint for extending a record and explain its request body.” The answer should cite the published reference and identify its source version.

Available tools

Tool	Use it to
search_docs	Find features, guides, examples, release notes and skills
read_doc	Read a document or one section
find_api_operations	Find endpoints by task, method, path or tag
get_api_operation	Inspect parameters, schemas, authentication and responses
list_skills	Discover supported tasks and target clients
read_skill	Read a skill and its supporting reference files

Skills are also available as MCP resources and downloadable bundles. Use the complete bundle when installing a skill so its relative references work.

Request limits

The public server allows 120 requests per IP address per minute. If you receive HTTP 429, wait a minute before retrying. Reuse retrieved documents while their source version has not changed.

Documentation and company access

The existing authenticated Caraer MCP is a separate service for working with company data. A skill can describe how to use it, but this developer MCP does not provide that access or authorise writes. Check the live company tool schema and permissions before acting on company data.

Updates

The MCP serves a validated snapshot of published docs. Results identify that snapshot’s version. If you cannot find a recently announced capability, check the changelog and the cited API reference instead of guessing an endpoint. A failed update leaves the previous validated snapshot available.