mcp-builder

# MCP Builder Create MCP servers that let agents interact with external services through clean tools. ## Process ### 1. Research - Review the MCP specification at `https://modelcontextprotocol.io/sitemap.xml`. - Fetch the TypeScript or Python SDK README from GitHub. - Read the target API docs to identify endpoints, auth, pagination, and data models. ### 2. Plan - List the tools to implement. Prefer comprehensive API coverage over one-off workflow tools. - Use consistent, action-oriented names with prefixes, e.g. `github_create_issue`, `github_list_repos`. - Decide transport: streamable HTTP for remote, stdio for local. ### 3. Implement **TypeScript (default)** - Use the MCP SDK with Zod for input schemas. - Define `outputSchema` and `structuredContent` where supported. - Annotate tools with `readOnlyHint`, `destructiveHint`, `idempotentHint`, `openWorldHint`. **Python** - Use FastMCP with Pydantic models. - Add `@mcp.tool()` decorators. Shared rules: - Async I/O for network calls. - Actionable error messages. - Pagination on list operations. - Return both text and structured data when possible. ### 4. Test - Build and run the server. - Test with the MCP Inspector: ```bash npx @modelcontextprotocol/inspector ``` ### 5. Evaluate Create 10 read-only, complex, realistic questions that require multiple tool calls. Verify answers yourself. Store as `evals.xml` or `evals/evals.json`. ## Reference - `references/mcp-best-practices.md` — server and tool conventions - `references/node-mcp-server.md` — TypeScript patterns - `references/python-mcp-server.md` — Python patterns - `references/evaluation.md` — evaluation guidelines