building-mcp-servers

Use the FastMCP framework from the official MCP Python SDK. Define tools with the @mcp.tool decorator, validate inputs with Pydantic models, set annotations like readOnlyHint, and run the server over stdio or streamable HTTP transport.

Use the MCP TypeScript SDK's McpServer class and registerTool method with Zod schemas for input validation. Provide explicit title, description, inputSchema, and annotations for each tool, then build with tsc and run via stdio or streamable HTTP.

Use stdio for local integrations and command-line tools running as subprocesses. Use streamable HTTP for remote servers, web services, and multi-client scenarios. SSE is deprecated in favor of streamable HTTP.

Create 10 read-only, complex evaluation questions in XML format with verifiable answers, then run the included evaluation.py script against your server. It reports accuracy, tool call counts, and agent feedback on tool quality.

FastMCP validates Host headers via transport security settings. Configure TransportSecuritySettings with allowed_hosts including your Docker container name and 0.0.0.0:* so requests from outside the container are accepted.

Do not use it when consuming an existing MCP server—connect to that server directly instead. It is intended only for creating new servers that integrate external APIs or services as LLM tools.