MCP Builder

What it does

Loads a structured four-phase workflow (research, implementation, review, evaluations) that Claude follows when building a Model Context Protocol server. It pulls in the official MCP spec, SDK READMEs, and language-specific reference files so Claude generates servers that match current protocol conventions — proper tool naming, Zod/Pydantic schemas, structured outputs, actionable errors, and pagination. It also forces Claude to produce a 10-question XML evaluation file at the end. Most useful when you’re starting a new MCP server from scratch or refactoring one that predates the modern SDK patterns.

Use cases

• Wrap an internal REST API as an MCP server so Claude Code or Desktop can call it directly — the skill guides tool selection, schema definition, and auth wiring in either TypeScript or Python.
• Refactor a hand-rolled MCP server to use structuredContent, outputSchema, and the tool annotation hints (readOnlyHint, destructiveHint, idempotentHint) that newer clients rely on.
• Ship an evaluation suite alongside a server — the skill’s Phase 4 produces a 10-question XML file with verifiable answers so you can regression-test whether an LLM can actually drive the tools you built.

When to use it

• You’re starting a new MCP server and want the current best-practice scaffolding instead of guessing from year-old blog posts.
• You already have a server but it’s missing structured outputs, pagination, or proper error messages, and clients are misusing the tools.
• You need an evaluation harness to prove a server works before shipping it to teammates.

When NOT to use it

• You just need to install and configure an existing MCP server — this skill builds servers, it doesn’t set up clients.
• You’re writing a one-off script or CLI; MCP adds protocol overhead that isn’t justified outside agent contexts.
• You need a stateful server with streaming responses — the skill steers toward stateless streamable HTTP and won’t design for long-lived sessions.
• Your target language isn’t Python or TypeScript; the reference guides only cover FastMCP and the TypeScript SDK.

Installation

/plugin install mcp-builder@anthropic-agent-skills

If you haven’t added the Anthropic marketplace yet, run this first:

/plugin marketplace add anthropics/skills

After install, just mention it in conversation:

Use the MCP Builder skill to create a TypeScript MCP server that wraps the Linear API and exposes tools for creating, listing, and updating issues.

Troubleshooting

Claude generates tools but skips the outputSchema and structured content. Prompt Claude to re-read the TypeScript or Python reference guide the skill loaded, and explicitly ask it to add outputSchema plus structuredContent returns to every tool. Older SDK examples in training data often omit these, so a direct instruction is needed.

The generated server uses stateful sessions or SSE transport. Tell Claude to follow the skill’s transport guidance: stateless streamable HTTP for remote servers, stdio for local. Regenerate the transport setup and remove any session management code.

Evaluation questions are trivial or answerable with one tool call. Re-run Phase 4 and require Claude to produce questions that need multiple tool calls, cross-referencing, or filtering. Reject any question whose answer is a direct field on a single API response.

npx @modelcontextprotocol/inspector can’t connect to the built server. Check that npm run build completed without errors and that the entry point in package.json points to the compiled JavaScript file, not the TypeScript source. For stdio servers, verify the shebang and executable bit on the built file.