Agent Skills: MCP (Model Context Protocol) Expert

MCP (Model Context Protocol) Expert

You are an expert in developing MCP servers for integration with Claude and other LLMs.

MCP Architecture

Core Concepts

• Tools: Functions that Claude can invoke (operations, queries, actions)
• Resources: Static or dynamic data exposed to Claude (files, templates)
• Prompts: Pre-defined prompt templates
• Transports: stdio (default), HTTP/SSE, WebSocket

SDK Setup (TypeScript)

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { z } from "zod";

const server = new McpServer({
  name: "my-mcp-server",
  version: "1.0.0",
});

Tool Structure

server.registerTool(
  "tool_name",
  {
    description: `Clear description of what the tool does.

When to use:
- Use case 1
- Use case 2

Limitations:
- Limitation 1`,
    inputSchema: z.object({
      param: z.string().describe("Parameter description"),
      limit: z.number().int().positive().max(1000).default(100)
        .describe("Max results (default: 100, max: 1000)"),
    }),
  },
  async (args) => {
    const result = await operation(args);
    return {
      content: [{ type: "text", text: JSON.stringify(result, null, 2) }],
      structuredContent: result,
    };
  }
);

Return Format

// Success
return {
  content: [{ type: "text", text: "Human readable result" }],
  structuredContent: { data: "Machine readable" },
};

// Error
return {
  content: [{ type: "text", text: `Error: ${error.message}` }],
  isError: true,
};

Best Practices

Security

1. Validate all inputs with Zod schemas
2. Use parameterized queries - never concatenate SQL
3. Principle of least privilege for database users
4. Rate limit expensive operations
5. Sanitize sensitive data in outputs

Tool Design

1. Single Responsibility: One tool = one specific operation
2. Descriptive Names: sql_select, file_read, not query or do
3. Clear Descriptions: Claude uses descriptions to decide when to use tools
4. Detailed Schemas: Use .describe() on every Zod field
5. Explicit Limits: maxRows, timeout, maxSize

Error Handling

try {
  const result = await db.query(sql);
  return { content: [{ type: "text", text: JSON.stringify(result) }] };
} catch (error) {
  console.error("[MCP] Error:", error);
  return {
    content: [{ type: "text", text: `Query failed: ${error.message}` }],
    isError: true,
  };
}

Logging

Important: Use console.error for logs in stdio servers (stdout is for protocol).

console.error("[MCP] Tool called:", toolName);
console.error("[MCP] Result:", JSON.stringify(result));

Claude Configuration

{
  "mcpServers": {
    "server-name": {
      "command": "node",
      "args": ["path/to/build/index.js"],
      "env": {
        "DB_HOST": "localhost",
        "DB_NAME": "mydb"
      }
    }
  }
}

Testing

# Use MCP Inspector for interactive testing
npx @modelcontextprotocol/inspector