MCP Server Development Guide

Overview

Create MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools.

High-Level Workflow

Phase 1: Deep Research and Planning

1. •Understand Modern MCP Design:
   • •Balance API coverage vs workflow tools.
   • •Use clear, descriptive tool names (e.g., github_create_issue).
   • •Provide concise tool descriptions and actionable error messages.
2. •Study Documentation:
   • •Protocol Specs: https://modelcontextprotocol.io/specification/draft.md
   • •Best Practices: [📋 View Best Practices](./reference/mcp_best_practices.md)
3. •Choose Stack:
   • •TypeScript (Recommended): High-quality SDK, strong typing.
   • •Python: Good for data/ML heavy integration.
4. •Plan Implementation:
   • •Review service API.
   • •Select tools to implement (common operations first).

Phase 2: Implementation

1. •Project Structure:
   • •TypeScript: package.json, tsconfig.json, src/index.ts.
   • •Python: pyproject.toml, server.py.
2. •Core Infrastructure:
   • •Auth, Error handling, Pagination.
3. •Implement Tools:
   • •Input Schema: Zod (TS) or Pydantic (Python).
   • •Output Schema: Structured data + text.
   • •Description: Summary, parameters, return type.
   • •Annotations: readOnlyHint, destructiveHint.

Phase 3: Review and Test

• •Code Quality: DRY, Error handling, Type coverage.
• •Build: npm run build or py_compile.
• •Test: Use npx @modelcontextprotocol/inspector.

Phase 4: Evaluations

• •Create 10 independent, read-only, complex questions to verify the server.
• •Format as XML evaluations.

Reference Files

(Note: Source references are available in external/skills_repos/anthropic-skills/skills/mcp-builder/reference)

• •Protocol Spec
• •TypeScript SDK
• •Python SDK