MCP Builder
Officialby Anthropic
MCP Builder is an official Anthropic agent skill that guides Claude Code through the full lifecycle of building production-quality Model Context Protocol (MCP) servers. It structures development across four phases: deep research and planning, implementation, review and testing, and evaluation creation. The skill helps you integrate any external API or service into the MCP ecosystem, covering every core MCP primitive — tools (with Zod or Pydantic schemas, structured output, and behavioral annotations), resources, and prompts. It supports both TypeScript (the recommended path, using the official MCP TypeScript SDK with Zod for input validation and streamable HTTP or stdio transport) and Python (using FastMCP with Pydantic models). Authentication patterns are covered as part of API client setup, including OAuth 2.1 and environment-variable-based API key management. The skill loads live SDK documentation, MCP best practices, and language-specific reference guides on demand during each phase. It concludes with a rigorous evaluation framework — generating 10 realistic, complex, read-only XML test questions to verify that an LLM can effectively use the finished server. Ideal for developers who want to expose any REST API, internal service, or data source to AI agents through a well-designed, production-grade MCP server.
Installation
Key Features
- ✓Four-phase structured workflow: research → implement → review/test → evaluate
- ✓TypeScript support with Zod schemas, registerTool/registerResource/registerPrompt patterns, and MCP Inspector testing
- ✓Python support via FastMCP with Pydantic model validation and decorator-based tool registration
- ✓On-demand loading of live MCP SDK documentation and best practices reference guides
- ✓Behavioral tool annotations (readOnlyHint, destructiveHint, idempotentHint) for agent discoverability
- ✓Automated evaluation generation: 10 complex XML question-answer pairs verifying real LLM task completion
Use Cases
- →Building a production MCP server that wraps a third-party REST API (Slack, GitHub, Stripe) for use in Claude Code or other MCP clients
- →Designing a local stdio MCP server to expose internal tools, databases, or file systems to AI agents
- →Creating a remote streamable HTTP MCP server for multi-client, stateless API integration deployments
- →Ensuring an existing MCP server meets quality standards: consistent naming, error handling, and type safety
- →Generating a comprehensive evaluation suite to benchmark how effectively an LLM uses a newly built MCP server