The mcp-builder skill from the anthropics/skills repository provides developers with a structured methodology to build Model Context Protocol (MCP) servers. The guide targets implementations where large language models interact with external services via tool calls. The documented workflow has four phases: research and planning, implementation, review and testing, and evaluation generation.
For project setup, the guide highlights TypeScript via the MCP SDK and Python via the Python SDK or FastMCP. TypeScript is recommended in the material for its static typing and environment support, pairing with streamable HTTP transports using stateless JSON for remote deployments, or stdio for local execution. Implementations define input schemas using Zod in TypeScript or Pydantic in Python. Depending on the external API being integrated, developers must supply appropriate API credentials and client authentication logic in their core infrastructure utilities.
Tool ergonomics focus on discoverability and context management. The guide advises action-oriented naming with uniform prefixes such as github_create_issue, concise functional descriptions, and actionable error messages that suggest concrete next steps. Responses should support pagination and filtering to keep context windows bounded, optionally providing structured outputs alongside standard text content. Review workflows recommend testing servers locally using @modelcontextprotocol/inspector.
To verify functionality, the guide specifies authoring an XML evaluation suite containing ten realistic, read-only questions. These questions require multi-tool reasoning, depend on non-destructive operations, and yield stable, verifiable single-string answers. The skill documentation is licensed under Apache-2.0.