MCP Builder

What is MCP

Model Context Protocol is the open standard for connecting AI agents to external tools and data sources. Think of it as USB-C for AI — one standard connector to any tool.

When to use MCP: When you want a tool to be reusable across multiple agents or platforms (Claude, Copilot, Cursor, etc.)

When NOT to use MCP: For project-specific, one-off integrations where the overhead isn't worth it.

MCP Server Structure

my-mcp-server/
├── server.py (Python) or index.ts (TypeScript)
├── tools/
│   ├── tool_one.py
│   └── tool_two.py
├── resources/       # Optional: expose data sources
├── prompts/         # Optional: expose prompt templates
├── requirements.txt or package.json
└── README.md

Python (FastMCP) — Recommended

python
from fastmcp import FastMCP

mcp = FastMCP("my-server-name")

@mcp.tool()
def get_user(user_id: str) -> dict:
    """
    Retrieve a user by their ID.
    
    Args:
        user_id: The unique identifier of the user
    
    Returns:
        User object with id, name, email
    """
    # implementation
    return {"id": user_id, "name": "James", "email": "james@example.com"}

@mcp.tool()
def create_task(title: str, description: str, assignee: str = None) -> dict:
    """Create a new task in the project management system."""
    # implementation
    return {"id": "task_123", "title": title, "status": "created"}

if __name__ == "__main__":
    mcp.run()

TypeScript (MCP SDK)

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-server", version: "1.0.0" })

server.tool(
  "get_user",
  "Retrieve a user by their ID",
  { userId: z.string().describe("The user's unique ID") },
  async ({ userId }) => ({
    content: [{ type: "text", text: JSON.stringify({ id: userId, name: "James" }) }]
  })
)

const transport = new StdioServerTransport()
await server.connect(transport)

Tool Design Principles

Tool Definition Checklist

• ] Name is descriptive and action-oriented (get_user, not user)
• ] Description clearly states when to use this tool
• ] Parameters have clear names and descriptions
• ] Returns are documented
• ] Error cases are handled and return meaningful messages
• ] Side effects are documented (if tool writes/deletes data)

Tool Design Rules

• One tool = one action — don't combine multiple operations
• Idempotent where possible — safe to call multiple times
• Validate inputs — don't trust that the AI passes valid data
• Return structured data — JSON is better than plain text for programmatic use
• Include context in errors — "User 'abc' not found" not just "Not found"

Resources (Exposing Data)

python
@mcp.resource("users://list")
def list_users() -> str:
    """Returns all users as a JSON string"""
    users = db.get_all_users()
    return json.dumps(users)

@mcp.resource("user://{user_id}")
def get_user_resource(user_id: str) -> str:
    """Returns a specific user's data"""
    user = db.get_user(user_id)
    return json.dumps(user)

Connecting to Claude Desktop

In claude_desktop_config.json:

json
{
  "mcpServers": {
    "my-server": {
      "command": "python",
      "args": ["/path/to/server.py"],
      "env": {
        "API_KEY": "your-key-here"
      }
    }
  }
}

Security Checklist

• ] Never expose credentials in tool definitions
• ] Validate and sanitize all inputs before using in DB queries or file operations
• ] Rate limit expensive operations
• ] Log all tool invocations for auditing
• ] Implement timeouts for external calls
• ] Don't give the tool more permissions than it needs