mcp-builder

Use FastMCP for Python or the official MCP TypeScript SDK with server.registerTool. Define input schemas with Pydantic or Zod, add tool annotations like readOnlyHint, and choose stdio transport for local servers or streamable HTTP for remote deployments.

MCP is a protocol that lets LLMs interact with external services through well-defined tools, resources, and prompts. Use it when you want AI assistants to call your API, query your data, or perform actions in external systems through a standardized interface.

Use stdio for local integrations and command-line tools where the server runs as a subprocess. Use streamable HTTP for remote servers, multi-client scenarios, and cloud deployments; SSE is deprecated in favor of streamable HTTP.

Create an XML evaluation file with 10 complex, read-only questions that have single verifiable answers, then run the evaluation.py script with your server connection. The harness uses Claude to attempt each question and reports accuracy, tool call counts, and agent feedback.

Use snake_case names with a service prefix like github_create_issue, and write descriptions that precisely state functionality, parameters, return schemas, and error cases. Include annotations such as readOnlyHint and destructiveHint so clients understand tool behavior.

Tools returning unbounded results exhaust the LLM's context window. Implement pagination with limit and offset parameters, enforce a character limit with truncation messages, and support filtering options so agents can request focused subsets of data.