xmcp

However much an AI chat assistant can explain, it cannot look up a single order in your shop's database until someone builds it a door. The Model Context Protocol (MCP) gives that door a standard shape: a small server that tells AI apps which actions and data an app offers. xmcp, an MIT-licensed TypeScript framework on GitHub, takes care of the protocol code for such servers, so developers write ordinary functions.

Each tool is one file in a folder, which xmcp finds and registers at startup. The file holds an input schema, a name, a description that tells the AI model when to use it, and the function that does the work. Prompts and resources, the read-only data a model can pull in, follow the same pattern.

One codebase can serve over HTTP for a hosted server or over STDIO, the local connection used by desktop apps such as Claude Desktop. An adapter adds the MCP server to an existing Next.js, Express, Fastify or NestJS app, so the server runs inside the app itself.

Middleware checks every request, for example for an API key or a rate limit. Plugins connect Clerk, Stripe and other services for sign-in and billing, and a template gallery offers starter servers to fork.