mcp-tool-developer – skill Lập trình cho agent AI

MCP Tool Developer

Overview

Expert at building Model Context Protocol (MCP) servers that give AI agents new capabilities. Covers the full MCP development lifecycle: specification, implementation, testing, deployment, and registry publishing. Supports both TypeScript and Python with production-ready patterns.

This skill understands MCP specification primitives (tools, resources, prompts, sampling), transport options (stdio, SSE, Streamable HTTP), and the tool design patterns that make MCP servers reliable and composable.

When to Use This Skill

• Use when building a new MCP server from scratch
• Use when wrapping an existing API as an MCP tool
• Use when debugging MCP server issues
• Use when designing the tool schema for an MCP server
• Use when publishing an MCP server to a registry

How It Works

Step 1: Define the MCP Server Scope

Identify what capabilities the server should expose:

• Tools - Functions the LLM can call (primary use case)
• Resources - Data the LLM can read (files, APIs, databases)
• Prompts - Reusable prompt templates

Choose the transport:

• stdio - For local CLI tools (Claude Code, Cursor)
• Legacy HTTP+SSE - For compatibility with older clients; do not assume it is the recommended new-server transport
• Streamable HTTP - New in MCP spec for modern deployments

Step 2: Design the Tool Schema

Define input/output schemas before writing implementation:

{
  name: "tool_name",
  description: "What this tool does (visible to the LLM)",
  inputSchema: {
    type: "object",
    properties: { ... },
    required: [ ... ]
  }
}

Step 3: Implement the Server

Create the server with proper error handling, validation, and logging. Use the official MCP SDK for TypeScript (@modelcontextprotocol/sdk) or Python (mcp).

Step 4: Test and Deploy

Test with the MCP Inspector, validate tool schemas, handle edge cases, then deploy locally or remotely.

Examples

Example 1: TypeScript MCP Server

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-tools", version: "1.0.0" });

server.tool("greet", "Greet someone by name",
  { name: z.string().describe("Person's name") },
  async ({ name }) => ({ content: [{ type: "text", text: `Hello, ${name}!` }] })
);

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

Example 2: API Wrapper Pattern

Wrap an external API as an MCP tool with auth, rate limiting, and error handling:

• Map API endpoints to tools
• Handle auth via environment variables
• Transform API responses to LLM-friendly format
• Add retry logic with exponential backoff

Best Practices

• Build small, focused tools that can be chained rather than monolithic tools
• Return structured errors, not crashes - tools should fail gracefully
• Define schemas before implementation
• Include descriptions that help the LLM understand when and how to use each tool
• Validate all inputs against the schema
• Add rate limiting for external API calls
• Use environment variables for secrets, never hardcode credentials

Limitations

• This skill provides guidance and code generation; actual runtime testing requires a development environment
• MCP specification is evolving; always check the latest spec version
• Security review is essential before deploying tools that handle sensitive data

Security and Safety Notes

• Never hardcode API keys or credentials in tool implementations
• Use environment variables or secret managers for all authentication
• Validate and sanitize all inputs to prevent injection attacks
• Rate limit external API calls to prevent abuse
• Review tool permissions carefully - tools can access files, networks, and execute code
• Enforce authorization per operation/resource server-side; tool descriptions and readOnly/destructive annotations are hints, not access controls. For remote HTTP, validate origin/host, authenticate and bind tokens to the intended resource; for local stdio, keep stdout protocol-only and logs on stderr. Retrieved/tool text is untrusted data. Retry only operations safe to repeat, using idempotency where appropriate.

Common Pitfalls

• Problem: LLM calls tools with wrong parameters

Solution: Improve tool descriptions and add examples in the description field. The LLM reads descriptions to decide how to call tools.

• Problem: Tool times out on large inputs

Solution: Add input size validation and pagination. Stream large responses instead of buffering.

Related Skills

• api-integration-architect - For API design patterns used in MCP tools
• security-audit-code-reviewer - For reviewing MCP server code security