fastmcp

Install fastmcp with pip, create a FastMCP instance, and decorate functions with @mcp.tool to expose them as tools. You can scaffold from the included api_wrapper, database_server, or file_processor templates, then validate with fastmcp inspect and fastmcp call.

Run fastmcp inspect server.py:mcp to verify the server loads, then fastmcp list to enumerate tools and fastmcp call with arguments to execute each tool. For HTTP transport, start with fastmcp run --transport http and call the local /mcp endpoint.

Yes, FastMCP supports direct installation with fastmcp install claude-code, claude-desktop, or cursor. Client installs run in isolated environments, so declare dependencies explicitly using flags like --with or editable installs.

Inspection fails when the file has import-time side effects that crash, the FastMCP object name does not match the file.py:object reference, or optional template dependencies are missing. Fix imports and install required packages first.

Use tools for actions, resources for stable read-only content like schemas or documents, and prompts for reusable LLM instruction templates. Start with tools only and add resources or prompts only when the client genuinely needs them.