Build an MCP server for Claude Code in 30 分

The piece is written from firsthand experience: the author had already built an MCP server for internal API integration and connected it to Claude Code, and describes the first thing that tripped them up as not implementation difficulty but the quiet failure caused by writing one console.log line. That failure mode follows from how a stdio server works — the process's stdout is the JSON-RPC channel itself, so anything else written there corrupts the conversation. The recommended logging escape hatch is console.error, which goes to stderr instead.

The recommended first tool is deliberately small: list_todos, which scans a directory for TODO and FIXME comments in .ts, .tsx, .js and .py files and returns them with file paths and line numbers. The author argues this kind of modest tool still matters, because it adds information a team already cares about into what the LLM can reach.

Beyond the stdout trap, the guide lists a few more practical snags. A loose inputSchema (the tool's argument definitions, written with the Zod library) makes the client's argument generation unstable, so the schema should be strict. The entry point passed to claude mcp add is safest as an absolute path, so the server does not depend on the directory Claude Code was launched from. And shutting down Claude Code without stopping the server can leave the child process lingering, which is why the author suggests checking with ps aux | grep tsx during development. The closing advice is to start with a minimal one-tool server and connect it via claude mcp add before expanding.