Baonat
← All skills
Developer workflow · ClaudeVerified

Author: Anthropic

Claude MCP Builder — Ship Model Context Protocol servers

Turn an API or internal service into a solid MCP server — planned tools, schemas, and evals — instead of a half-wired stdio script that breaks on the first bad payload.

What it does
Guides Claude through research → implement → review → evaluate for MCP servers (Python FastMCP or TypeScript MCP SDK), with tool naming, annotations, and realistic eval questions.
Best for
Developers wiring Claude/Cursor to external APIs, internal admin tools, or data services via Model Context Protocol.
How to install / use
Clone or enable mcp-builder from anthropics/skills (example-skills / Apache-2.0). Provide the target API docs URL, language choice, and transport (stdio vs HTTP) before implementation.

Attach mcp-builder before the first scaffolding commit — not after the messy prototype.

Copy this

  • Pick language + transport up front: TypeScript SDK + stdio/HTTP
  • Require tool annotations (readOnly / destructive) and JSON schemas
  • Ship ≥10 eval questions with verifiable, preferably read-only answers

When to use

Use Anthropic’s mcp-builder skill when you are creating or hardening an MCP server so Claude (or another MCP client) can call a real external API or internal service through typed tools — not when you only need a one-off curl from a chat thread.

Good fits:

  • Wrapping a SaaS API (tickets, CRM, billing) as MCP tools for agents.
  • Exposing internal admin/read APIs to Cursor or Claude Desktop.
  • Reworking a prototype server into schemas, annotations, and eval coverage.
  • Choosing Python (FastMCP) vs TypeScript (MCP SDK) with a deliberate transport.

Skip it for plain REST client code with no MCP, or for end-user document skills (use pdf/docx/xlsx/pptx instead).

Why

MCP servers fail in boring ways: vague tool names, missing schemas, destructive tools without hints, zero evals. Anthropic’s public mcp-builder skill (in anthropics/skills) encodes a phased build loop so the agent plans coverage, implements carefully, and tests with Inspector-style checks.

Baonat marks this verified. The skill lives under Anthropic’s example-skills set and is Apache-2.0 — free to clone and adapt. Host Claude/Cursor usage may still require your normal product subscription.

Install

  1. Get the skill from skills/mcp-builder, or install Anthropic’s example-skills plugin path where available (/plugin install example-skills@anthropic-agent-skills in Claude Code).
  2. Provide API context: docs URL, auth model, must-have vs nice-to-have operations.
  3. Choose stack: language (TypeScript recommended for broad SDK support) + transport (stdio local vs Streamable HTTP remote).
  4. Run the phased loop: research/plan → implement tools + schemas → review/test → write evals.
  5. Verify with MCP Inspector (or your client) on read-only tools first.
  6. Gate destructive tools behind explicit confirmation patterns.

Copy this

Kickoff prompt:

Attach the mcp-builder skill. Target API: [docs URL]. Auth: [scheme].
Language: [TypeScript MCP SDK | Python FastMCP]. Transport: [stdio | Streamable HTTP].
Must-have tools: [list]. Out of scope: [list].
Phase 1 ONLY: plan tool list with names, readOnly/destructive hints, and input schemas — no code yet.
After approval: implement, then add ≥10 eval questions with verifiable answers (prefer read-only).
Do not commit secrets; use env vars for credentials.

Concrete scenarios

  1. Support agent. MCP tools for “search tickets / post internal note / fetch customer plan.”
  2. Internal metrics. Read-only tools over a metrics API with evals that check known dashboards.
  3. Prototype cleanup. Rename tools, add schemas/annotations, and drop untested write endpoints.

Pitfalls / when NOT to use

  • Write tools first. Start read-only; add mutations after evals pass.
  • Secret leakage. Never bake tokens into SKILL outputs or repo files.
  • Kitchen-sink APIs. Prefer a small high-value tool set over wrapping every endpoint.

FAQ

Q: License?
A: Apache-2.0 example skill — free to clone from anthropics/skills.

Q: Official Anthropic skill?
A: Yes — mcp-builder under the public skills repo / example-skills plugin.

Q: Do I need Claude paid?
A: The GitHub artifact is free. Running the build inside Claude Code / Claude still follows your host plan.

Source attribution