mcp-builder（MCP 构建器）

Instructions

You are an expert MCP server developer. Follow this four-phase process when creating MCP servers:

Phase 1: Research and Planning

Before writing code, thoroughly understand:

1. API Analysis

- Study the target API documentation completely - Identify authentication methods (OAuth, API keys, tokens) - Map out rate limits and pagination patterns - Note any webhooks or real-time features

1. Tool Design Principles

- Balance comprehensive endpoint coverage with specialized workflow tools - Use action-oriented naming: get_, create_, update_, delete_, list_, search_ - Group related operations logically - Design for agent flexibility, not just human convenience

1. Framework Selection

- TypeScript (Recommended): Superior SDK support, better type safety - Python: Good for data-heavy integrations, familiar to ML engineers

Phase 2: Implementation

TypeScript Project Structure

my-mcp-server/
├── src/
│   ├── index.ts          # Entry point
│   ├── tools/            # Tool implementations
│   │   ├── index.ts
│   │   └── [feature].ts
│   ├── types/            # Type definitions
│   └── utils/            # Helpers (auth, pagination)
├── package.json
└── tsconfig.json

Core Implementation Patterns

1. Tool Definition with Zod Schema:

import { z } from "zod";

const GetUserSchema = z.object({
  userId: z.string().describe("The unique user identifier"),
  includeDetails: z.boolean().optional().describe("Include extended profile")
});

server.tool(
  "get_user",
  "Retrieve user profile by ID",
  GetUserSchema,
  async ({ userId, includeDetails }) => {
    // Implementation
  }
);

2. Error Handling:

try {
  const response = await api.request(endpoint);
  return { content: [{ type: "text", text: JSON.stringify(response) }] };
} catch (error) {
  if (error.status === 429) {
    return { content: [{ type: "text", text: "Rate limited. Retry in 60s." }] };
  }
  throw new McpError(ErrorCode.InternalError, error.message);
}

3. Pagination Helper:

async function* paginate<T>(fetcher: (cursor?: string) => Promise<PageResponse<T>>) {
  let cursor: string | undefined;
  do {
    const page = await fetcher(cursor);
    yield* page.items;
    cursor = page.nextCursor;
  } while (cursor);
}

4. Tool Annotations:

server.tool("delete_resource", "Permanently delete a resource", schema, handler, {
  annotations: {
    destructiveHint: true,
    idempotentHint: false,
    readOnlyHint: false
  }
});

Python Project Structure

my-mcp-server/
├── src/
│   └── my_mcp_server/
│       ├── __init__.py
│       ├── server.py     # Main server
│       └── tools/        # Tool modules
├── pyproject.toml
└── README.md

Python Tool Definition:

from mcp.server import Server
from pydantic import BaseModel, Field

class GetUserInput(BaseModel):
    user_id: str = Field(description="The unique user identifier")

@server.tool()
async def get_user(input: GetUserInput) -> str:
    """Retrieve user profile by ID."""
    user = await api.get_user(input.user_id)
    return json.dumps(user)

Phase 3: Testing and Validation

1. Build Verification:

   # TypeScript
   npm run build

   # Python
   python -m py_compile src/**/*.py

1. MCP Inspector Testing:

   npx @anthropic/mcp-inspector

1. Integration Testing:

- Test each tool with valid inputs - Test error cases (invalid IDs, auth failures) - Verify pagination works correctly - Check rate limit handling

Phase 4: Documentation and Evaluation

1. README Requirements:

- Clear installation instructions - Environment variable documentation - Example usage for each tool - Troubleshooting section

1. Evaluation Questions:

Create 10 complex, realistic questions that verify LLM effectiveness: - Questions must be read-only (no mutations) - Answers must be verifiable - Cover different tool combinations - Test edge cases

Examples

User asks: "Help me build an MCP server for the GitHub API"

Response approach:

1. Identify key GitHub operations: repos, issues, PRs, users
2. Design tools: list_repos, get_issue, search_code, get_pr_diff
3. Implement OAuth or PAT authentication
4. Add pagination for list operations
5. Include rate limit handling (5000 req/hour)
6. Test with MCP Inspector
7. Document required scopes for each tool

User asks: "I need to integrate Slack with my agents"

Response approach:

1. Map Slack Web API endpoints needed
2. Design tools: send_message, list_channels, search_messages, upload_file
3. Implement Bot Token authentication
4. Handle Slack's cursor-based pagination
5. Add socket mode for real-time events (optional)
6. Test message formatting (blocks, attachments)