MCP Builder

Overview

Anthropic's official guide for building MCP (Model Context Protocol) servers. Create high-quality MCP servers that enable LLMs to interact effectively with external services.

Four-Phase Workflow

Phase 1: Deep Research and Planning

1. Understand modern MCP design

   • Study MCP protocol docs and best practices
   • Understand Tools vs. Resources vs. Prompts
   • Learn Schema design and Annotations specification

2. Research the target service API

   • Read API documentation thoroughly
   • Understand authentication flows and rate limits
   • Identify high-value vs. low-value operations

3. Plan implementation

   • Decide tool granularity: API-coverage vs. workflow-oriented
   • Design tool names and parameter schemas
   • Plan error handling strategy

Phase 2: Implementation

1. Project structure

   • Choose SDK (TypeScript or Python)
   • Set up standard project skeleton
   • Configure type checking and linting

2. Core infrastructure

   • Implement authentication and session management
   • Set up logging and error reporting
   • Configure rate limit handling

3. Tool implementation

   • Every tool needs clear JSON Schema
   • Use Annotations to describe side effects
   • Strict input validation, consistent output format

Phase 3: Review and Test

• Code quality review
• Interactive testing with MCP Inspector
• Verify all tools under boundary conditions
• Ensure error messages are helpful to LLMs

Phase 4: Create Evaluations

Write 10 evaluation questions testing LLM tool usage:

• Cover simple and complex scenarios
• Include multi-step workflows
• Test error handling paths

Tool Design Principles

Reference Resources

• MCP Best Practices Guide
• TypeScript SDK / Python SDK documentation
• MCP Evaluation Guide
• Tool Annotations specification