You want to build an MCP server in Python. What do you install and what does the API look like?

The official **Python MCP SDK** is the `mcp` package on PyPI. You install it with `pip install mcp` (or `uv add mcp` if you use uv). The source lives at github.com/modelcontextprotocol/python-sdk and is maintained by the protocol team. The SDK's main feature is FastMCP, a decorator-based API styled after FastAPI. A working tool server looks like this: ```python from mcp.server.fastmcp import FastMCP app = FastMCP("my-server") @app.tool() def add(a: int, b: int) -> int: """Add two numbers.""" return a + b if __name__ == "__main__": app.run() ``` That is the complete file. The SDK reads the type hints on `add` and generates a JSON Schema describing the arguments. It uses the docstring as the tool description the LLM sees. All JSON-RPC framing, the initialize handshake, the tools/list and tools/call methods, and lifecycle handling happen inside `app.run()`. The SDK ships with stdio and streamable HTTP transport adapters. Stdio is the default. Point the file path in Claude Desktop's config and the add tool appears in the chat surface. The same decorator pattern extends to `@app.resource()` for readable URIs and `@app.prompt()` for templated prompts, covering all three MCP primitives.

The official **Python MCP SDK** is the `mcp` package on PyPI (`pip install mcp`), maintained at github.com/modelcontextprotocol/python-sdk by the protocol team. Python is overrepresented in ML and data tooling, which makes this the most popular SDK for internal MCP servers. The SDK ships FastMCP, a decorator-driven API borrowing its shape from FastAPI: ```python from mcp.server.fastmcp import FastMCP app = FastMCP("my-server") @app.tool() def add(a: int, b: int) -> int: """Add two numbers.""" return a + b if __name__ == "__main__": app.run() ``` Three things make this concise. First, JSON Schema for tool arguments is auto-generated from Python type hints (Pydantic models work too), so you never hand-write schemas. Second, the docstring becomes the tool description that the LLM sees, keeping documentation in code. Third, all JSON-RPC framing, the initialize handshake, and lifecycle handling are inside `app.run()`. Transports ship in the box. `app.run()` defaults to stdio, which is the transport Claude Desktop, Cursor, and Zed all use for local servers. Pass `transport="streamable-http"` and the same server speaks the remote transport instead. The decorator surface extends beyond tools to `@app.resource()` (readable URIs) and `@app.prompt()` (templated prompt starters), covering all three MCP primitives in one consistent API. Below FastMCP sits a low-level Server API (`mcp.server.lowlevel.Server`) for advanced cases like custom transports or session multiplexing. FastMCP is built on top of it. New code should default to FastMCP and only descend when hitting a genuine wall.