MCP Builder

The MCP Builder skill is the official Anthropic Agent Skill that guides agents through creating high-quality MCP (Model Context Protocol) servers — the connectors that let LLMs interact with external services through well-designed tools, in either Python (FastMCP) or Node/TypeScript (MCP SDK).

What is the MCP Builder skill?

It is a skill from Anthropic's anthropics/skills repository that turns the job of writing an MCP server into a structured, four-phase workflow: deep research and planning, implementation, review and testing, and evaluation. It triggers whenever you set out to build an MCP server that integrates an external API or service. The quality bar it enforces is concrete — a server is judged by how well it actually enables an LLM to complete real-world tasks, not by raw endpoint count.

What it does

• Teaches modern MCP server design: API coverage versus workflow tools, clear action-oriented tool naming with consistent prefixes (e.g. github_create_issue), context management, and actionable error messages.
• Points you to the MCP specification (via modelcontextprotocol.io/sitemap.xml and .md pages) and the official Python SDK and TypeScript SDK READMEs.
• Provides language-specific implementation guides for FastMCP (Python, with Pydantic models and the @mcp.tool decorator) and the TypeScript MCP SDK (with Zod schemas and server.registerTool).
• Recommends transports — streamable HTTP with stateless JSON for remote servers, stdio for local — and tool annotations (readOnlyHint, destructiveHint, idempotentHint, openWorldHint).
• Drives a review pass for DRY code, full type coverage, and clear descriptions, then build/test with MCP Inspector (npx @modelcontextprotocol/inspector).
• Generates a 10-question evaluation suite in XML to prove the server lets an LLM answer realistic, complex, verifiable questions.

How it works

The skill loads a documentation library on demand. Phase 1 has you read the MCP best-practices reference and the relevant SDK README; Phase 2 sets up project structure, shared infrastructure (API client, error handling, response formatting, pagination), and per-tool input/output schemas; Phase 3 reviews code quality and runs the build plus MCP Inspector; Phase 4 creates evaluations following the evaluation guide. Reference files (reference/mcp_best_practices.md, reference/python_mcp_server.md, reference/node_mcp_server.md, reference/evaluation.md) are pulled in only when each phase needs them.

Who it is for

Developers building MCP servers to connect Claude or other LLMs to external APIs and internal services — platform engineers wrapping a SaaS API, teams exposing internal tooling to agents, and anyone shipping a server in Python with FastMCP or in Node/TypeScript with the MCP SDK. TypeScript is the recommended default for its SDK support and tooling.

What you can build

Production MCP servers that wrap third-party APIs (GitHub, Slack, internal services), workflow tools that compose multiple endpoints into one agent-friendly call, remote servers over streamable HTTP, local stdio servers, and a repeatable evaluation harness that scores how well an LLM uses your tools.

Why it matters

Most hand-written MCP servers fail agents quietly — vague tool names, oversized responses that blow the context window, and unhelpful error messages. This skill encodes Anthropic's own best practices for tool design, schemas, transports, and evaluation, so the servers your agent builds are discoverable, context-efficient, and provably usable by an LLM.

1. Install via the skills CLI

npx skills add anthropics/skills --skill mcp-builder

2. Or add it as a Claude Code plugin

/plugin marketplace add anthropics/skills
/plugin install example-skills@anthropic-agent-skills

3. Use it

Ask your agent to build an MCP server for a specific API — for example, "build an MCP server for the GitHub API in TypeScript" — and the skill walks through research, implementation, testing with MCP Inspector, and evaluation.