MCP Builder skill

MCP Builder is the skill to turn on when the tool you want Claude to use has no connector yet. It gives Claude a four-phase plan for writing an MCP server, a small program that exposes another service’s actions as tools Claude can call. MCP stands for Model Context Protocol, the open standard behind Claude’s connectors.

What it does (plain English)

The skill is a development guide plus reference files Claude loads at the right moment. Its phases:

Phase 1: Research and plan. Claude studies the MCP specification and the official SDK documentation, then reads the target service’s API docs to list the endpoints worth exposing. The skill’s design advice: name tools clearly with a shared prefix (its examples are github_create_issue and github_list_repos), keep responses focused with filtering and pagination, and write error messages that tell the agent what to try next.

Phase 2: Implement. It recommends TypeScript, with Streamable HTTP for servers you host remotely and stdio for servers that run on your own computer. Python with FastMCP is the documented second option. Each tool gets an input schema (Zod for TypeScript, Pydantic for Python), a clear description, and annotations that say whether it only reads, whether it can destroy data, and whether it reaches the open internet.

Phase 3: Review and test. Build the project and test it with the MCP Inspector (npx @modelcontextprotocol/inspector).

Phase 4: Evaluate. Claude writes 10 realistic, read-only questions that need several tool calls, solves each one to confirm the answer, and saves them in an XML format. The folder includes a Python script to run these evaluations.

File in the skill	What it holds
SKILL.md	The four-phase workflow
reference/mcp_best_practices.md	Core design guidelines
reference/node_mcp_server.md	TypeScript patterns and examples
reference/python_mcp_server.md	Python patterns and examples
reference/evaluation.md	How to write evaluation questions
scripts/evaluation.py	Runs the evaluations (needs anthropic and mcp Python packages)

Is it for you?

This is the most technical skill on VibeAtlas. It suits you if you are comfortable running commands in Claude Code and you have a specific service in mind, such as your company’s internal API, a niche SaaS tool or a database Claude cannot reach.

If a connector already exists in the Connectors directory, use that instead. A vendor-maintained connector will be kept up to date; one you build is yours to maintain.

What you can do with it

1. Wrap an API. “Use the mcp-builder skill to build an MCP server for our bookings API. The OpenAPI file is in docs/api.yaml. Start with read-only tools.”
2. Review a server. “Check my MCP server’s tool names, descriptions and annotations against the mcp-builder guidelines.”
3. Write evaluations. “Create the 10 evaluation questions for my server and run them.”
4. Choose transport. “This server will run on my laptop only. Set it up with stdio.” Then add it to Claude Code with claude mcp add.

How to set it up

Claude Code

1. Run /plugin marketplace add anthropics/skills
2. Run /plugin install example-skills@anthropic-agent-skills. That bundle holds 12 skills; Claude Code loads each one only when it matches, and estimates mcp-builder at about 100 tokens always loaded and about 3k tokens when used.
3. To install only this skill, copy the skills/mcp-builder folder from GitHub to ~/.claude/skills/mcp-builder/ (the Apache 2.0 license allows it).
4. Ask Claude to build your server, then register it, for example claude mcp add my-server -- node build/index.js.

We ran the marketplace and install commands in Claude Code 2.1.302 on 2026-10-18; both succeeded. We did not build a server with the skill.

claude.ai or the desktop app

1. Turn on Code execution and file creation in Settings > Capabilities.
2. Look in Customize > Skills; the anthropics/skills README says its examples are already available on paid plans.
3. If it is not listed, zip the mcp-builder folder and upload it there.

You can plan and write the code in claude.ai, but you will run and host the server somewhere else.

Watch out for

• You own the security. A server holds API keys and can take actions. Start read-only and add write tools deliberately, with the destructive annotation set.
• Remote servers need hosting and auth. Connecting a remote server to claude.ai as a custom connector means running it on the internet with proper login. Anthropic’s custom connectors are not verified by Anthropic.
• Evaluations call the API. The evaluation script uses the anthropic Python package, so it needs an Anthropic API key and is billed to that account.
• Repo disclaimer. The anthropics/skills README calls these skills demonstrations; test before relying on them.

Alternatives

• Skill Creator when Claude just needs instructions, not a new connection.
• Fetch MCP server or Filesystem MCP server, small reference servers worth reading before you write your own.
• Zapier MCP to reach many apps without writing code.