The measure of an MCP server is not "it works" — it is whether the LLM gets real work done through it. Anthropic's official MCP Builder Skill is a complete MCP server development guide, covering all four phases: research and planning, protocol study, framework selection, and construction — FastMCP for Python, MCP SDK for Node/TypeScript.
The four-phase build flow
Phase 1: Research & planning — balance comprehensive API coverage with workflow tools; name tools clearly with prefixes (github_create_issue); keep descriptions concise with pagination/filtering; write error messages that guide agents toward solutions
Phase 2: Study the MCP protocol — start from the spec: architecture, transports (streamable HTTP/stdio), and tool/resource/prompt definitions, then the framework docs
Phase 3: Choose a framework — FastMCP for a quick Python start, the official MCP SDK for Node/TypeScript
Phase 4: Build & polish — implement against the design principles above, repeatedly validating tool descriptions and error-handling quality
Design principles: design tools for the LLM
The skill keeps returning to one lens: tools are for the LLM, not the human. So naming must be discoverable (consistent verb prefixes), context must be lean (return focused, relevant data — no full dumps), and errors must be actionable (tell the agent what to do next). Performance varies by client: some clients benefit from code execution that composes basic tools, others prefer higher-level workflow tools — when in doubt, prioritize comprehensive API coverage.
Typical use cases
External service integration: wrap company APIs as MCP tools so agents can call them directly (ticketing, CRM, databases)
Custom tool sets: build business-specific MCP servers mixing workflow tools with base API coverage
Agent capability expansion: mount your team's internal services onto coding agents for a closed business loop
How to use & download
Claude Code users: drop skills/mcp-builder into your project's .claude/skills/ folder. Describe the MCP server you need and the agent runs the four-phase flow. Full source: official GitHub repo.
Other agent users: follows the standard SKILL.md convention. Official docs: Anthropic Agent Skills documentation.
Related resources
MCP skills work in pairs: mcporter calls the servers you build; ACP Router (acp-router) routes coding-agent requests to the right environment.