mcp-builder

Use the official MCP TypeScript SDK's McpServer class and register tools with server.registerTool, providing title, description, Zod inputSchema, and annotations. Build with npm run build and test using the MCP Inspector via npx @modelcontextprotocol/inspector.

Use FastMCP from the MCP Python SDK, defining tools with the @mcp.tool decorator and Pydantic models for input validation. Verify syntax with python -m py_compile and test the server with the MCP Inspector.

Use streamable HTTP for remote servers serving multiple clients, and stdio for local integrations and command-line tools. SSE is deprecated in favor of streamable HTTP, and stdio servers must log to stderr rather than stdout.

Create an XML file with 10 read-only, complex QA pairs, then run the evaluation.py script with your transport type, server command or URL, and an Anthropic API key. The harness reports accuracy, tool call counts, and agent feedback per task.

Use snake_case names with a service prefix like github_create_issue, starting with action verbs. Descriptions must precisely match functionality, document parameters and return schemas, and include annotations such as readOnlyHint and destructiveHint.

Answers must be verifiable by direct string comparison, so questions must target historical or closed data that will not change over time. Questions about current state, like open issue counts, produce unreliable evaluation results.