MCP Builder

MCP Builder

Guide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools.

Prerequisites

• TypeScript (recommended) or Python
• Node.js 18+ for TypeScript, Python 3.10+ for Python
• Understanding of async/await patterns
• Familiarity with API integration

Instructions

1. Phase 1: Research and Planning

   • Study MCP Protocol: Start with https://modelcontextprotocol.io/sitemap.xml
   • Understand the API you're integrating
   • Plan tool selection prioritizing comprehensive API coverage

2. Phase 2: Set Up Project

   For TypeScript (recommended):

   npm init -y
   npm install @modelcontextprotocol/sdk zod
   

   For Python:

   pip install mcp pydantic
   

3. Implement Core Infrastructure

   • API client with authentication
   • Error handling helpers with actionable messages
   • Response formatting (JSON/Markdown)
   • Pagination support

4. Implement Tools

   Each tool needs:

   • Input Schema: Use Zod (TypeScript) or Pydantic (Python)
   • Output Schema: Define outputSchema for structured data
   • Clear Description: Concise summary with parameter descriptions
   • Annotations: readOnlyHint, destructiveHint, idempotentHint

   server.registerTool({
     name: "github_create_issue",
     description: "Create a new issue in a GitHub repository",
     inputSchema: z.object({
       repo: z.string().describe("Repository in owner/name format"),
       title: z.string().describe("Issue title"),
       body: z.string().optional()
     }),
     annotations: { destructiveHint: false, idempotentHint: false }
   });
   

5. Phase 3: Test

   # TypeScript
   npm run build
   npx @modelcontextprotocol/inspector
   
   # Python
   python -m py_compile your_server.py
   

6. Phase 4: Create Evaluations

   • Create 10 complex, realistic test questions
   • Each question should require multiple tool calls
   • Verify answers are stable and verifiable

Error Handling

• Return actionable error messages that guide toward solutions
• Include specific suggestions and next steps
• Handle rate limiting gracefully

Notes

• Use consistent tool naming prefixes (e.g., github_create_issue)
• Balance API coverage with workflow convenience tools
• Transport: Streamable HTTP for remote, stdio for local
• Return both text content and structured data

Source: anthropics/skills