Skip to content

Kelvin MCP server overview

The Kelvin MCP server exposes the full Kelvin API to MCP-compatible AI coding tools such as Claude, Cursor, and VS Code. It proxies every tool call directly to the Kelvin API using your authenticated session, with no custom integration required.

How it generates tools

On startup, the server reads the Kelvin API's OpenAPI specification and builds an MCP tool for each operation automatically.

  • One tool per operation, named after its operation ID (getAsset, listWorkloads, createRecommendation).
  • Each tool's input schema is derived from the endpoint's path and query parameters and request body.
  • Deprecated operations and non-JSON endpoints are excluded.

A built-in health_check tool is also available and returns the authenticated user identity and server version.

Because the catalog is generated from the spec, it mirrors the current Kelvin API surface without manual maintenance.

Request lifecycle

Clients communicate with the server over Streamable HTTP (JSON-RPC 2.0) at the /mcp endpoint.

The three MCP methods are:

  1. initialize: Handshake that returns server info and capabilities.
  2. tools/list: Returns the generated tool catalog.
  3. tools/call: The server builds the upstream request (filling path and query parameters, encoding the body), calls the Kelvin API with your bearer token, and returns the response.

The proxy applies gzip compression and caps response size to prevent large payloads from flooding the assistant's context window.

Next steps