mcp-builder

Use the MCP Python SDK with FastMCP to register tools via decorators, define Pydantic v2 models for input validation, and use async/await for all I/O. The Skill's Python implementation guide covers server initialization, tool registration, and a quality checklist.

Use the MCP TypeScript SDK with server.registerTool, define Zod schemas with .strict() for validation, enable TypeScript strict mode, and compile with npm run build. The Skill's TypeScript guide provides project structure and complete working examples.

MCP servers are long-running processes waiting on stdio or HTTP, so running them directly hangs your terminal. Use the evaluation harness, run the server in tmux, or use a timeout like 'timeout 5s python server.py'.

Good evaluation questions are read-only, independent, complex enough to require multiple tool calls, realistic, and have a single stable verifiable answer. Avoid questions about current state that changes over time or ones solvable by simple keyword search.

Use stdio for local command-line tools and desktop integrations, HTTP for remote multi-client web services, and SSE when clients need real-time server-push updates. The evaluation script supports all three transports.

Tools should be designed around agent workflows, not API structure—consolidate related operations into high-impact tools, return concise high-signal responses, and write actionable error messages that guide the LLM toward correct usage.