MCP Server Development Guide

Create MCP servers whose quality is measured by how well they let an LLM accomplish real tasks — not by raw endpoint count.

Workflow

1. Research & plan — balance broad API coverage against a few high-leverage workflow tools; design tool names, inputs, and outputs for the model, not the API.
2. Implement — FastMCP (Python) or the MCP SDK (Node/TS). Keep tool schemas tight; return structured, model-readable results.
3. Evaluate — write task-level evals and iterate on tool ergonomics.

Principles

• Prefer fewer, composable tools with clear contracts.
• Make errors actionable (tell the model what to do next).
• Document side effects and idempotency.