Turn Any Python Function Into a Tool Claude Code Can Call: Your First MCP Server in 15 Lines

TL;DR: The MCP Python SDK turns any typed Python function into a Claude-callable tool in 15 lines — decorators handle the JSON schema, the CLI handles the inspector, no protocol code needed.

What you'll be able to do after this: expose your own tools, APIs, or data sources to Claude Code (and any MCP host) so they're available as first-class tools in every session — without touching protocol or serialization code.

• @mcp.tool() is all you write. Type hints become JSON schema automatically; the SDK handles request parsing, validation, and transport.
• mcp dev server.py launches a browser inspector so you can invoke your tools and inspect the protocol messages before connecting anything to Claude.
• Use MCP when multiple agents or editors need the same capability. For a single session's private helper, a shell tool is simpler. MCP pays off when the tool is shared.

Minimal server (Python, 15 lines):

from mcp.server import MCPServer

mcp = MCPServer("my-tools")

@mcp.tool()
def search_docs(query: str, max_results: int = 10) -> list[dict]:
    """Search the internal docs API and return matching pages."""
    return fetch_docs(query, limit=max_results)

if __name__ == "__main__":
    mcp.run()

Steps:

1. Install: pip install "mcp[cli]" (Python 3.10+; or uv add "mcp[cli]")
2. Decorate each capability with @mcp.tool(), @mcp.resource(), or @mcp.prompt()
3. Test: mcp dev server.py — opens MCP Server Inspector in your browser; invoke tools and inspect JSON without connecting Claude
4. Wire to Claude Code — add to ~/.claude/settings.json:

{
  "mcpServers": {
    "my-tools": { "command": "python", "args": ["path/to/server.py"] }
  }
}

Then run claude mcp login my-tools if the server requires authentication, and restart Claude Code.

Three things that don't work automatically:

• Authentication: MCP has no built-in auth for tool calls. Wire a shared secret or OAuth yourself, or use the claude mcp login interactive flow.
• Error propagation: unhandled exceptions become opaque errors. Return a structured dict with an error key for failures the model can read and act on.
• Response size: the SDK serializes anything you return, but Claude Code enforces a per-tool result-size limit. Keep responses under ~50K characters or paginate.

The SDK is at v2, aligning with the 2026-07-28 MCP specification, which changes transports (Streamable HTTP is the new default; SSE still works). If you have a v1 server, check the migration guide — the API surface changed.

Start here: the official MCP quickstart at modelcontextprotocol.io builds a two-tool weather server end-to-end and connects it to Claude Desktop. Anthropic's free course Introduction to Model Context Protocol goes deeper — all three primitives (tools, resources, prompts) and a document management system you build and test with the MCP Server Inspector.

Sources: Build an MCP server — modelcontextprotocol.io · MCP Python SDK — GitHub · Introduction to MCP — Anthropic Skilljar · Step-by-step guide — Composio