mcp_builder

Install fastmcp with pip, instantiate FastMCP with a server name, and register functions using the @mcp.tool() decorator. Call mcp.run() to start the server, and add resources with @mcp.resource() for read-only data.

Install @modelcontextprotocol/sdk via npm, create an McpServer instance with name and version, and register tools with server.tool(). Connect it using StdioServerTransport so MCP clients can communicate over standard input and output.

FastMCP suits Python developers who want decorator-based tool definitions with minimal boilerplate. The TypeScript SDK fits Node.js environments and teams already working in the JavaScript ecosystem.

Add an entry to mcp_config.json under mcpServers with the command, args pointing to your server script, and any required env variables like API keys. Restart the client so it launches the server process.

Use MCP Inspector to manually call tools and verify endpoints return correct results. Then run an end-to-end integration test with the target client such as Claude Desktop to confirm tool invocation works.

Restrict file system access to allowed directories, sanitize user inputs to prevent path traversal, and validate parameters against JSON Schema. Return meaningful error messages so the LLM can handle failures gracefully.