Как построить AI Agent с MCP
MCP (Model Context Protocol) от Anthropic — standard для exposing tools к LLM agents. Server implements tools → Client (Claude Desktop, Zed, custom) connects → LLM invokes tools через structured calls. Simpler чем custom function calling. Implementations: Python @modelcontextprotocol/sdk, TypeScript, Go. Use cases: file system access, GitHub operations, database query, custom APIs.
Ниже: пошаговая инструкция, рабочие примеры, типичные ошибки, FAQ.
Пошаговая настройка
- Install MCP SDK:
pip install mcpилиnpm install @modelcontextprotocol/sdk - Define tools: name, description, input schema (JSON Schema)
- Implement handler: для каждого tool функция, возвращающая result
- Start MCP server: stdio transport (для local) или SSE (для remote)
- Connect client: Claude Desktop config, Zed settings, или custom agent
- Test: LLM автоматически разрешит available tools
- Deploy: self-host server или публикуйте как open-source
Рабочие примеры
| Сценарий | Конфиг |
|---|---|
| Python MCP server | from mcp.server.fastmcp import FastMCP
mcp = FastMCP('Enterno MCP')
@mcp.tool()
def check_dns(domain: str) -> dict:
"""Check DNS records for a domain."""
return {'a': resolve_a(domain), 'mx': resolve_mx(domain)}
if __name__ == '__main__':
mcp.run() |
| Claude Desktop config | # ~/Library/Application Support/Claude/claude_desktop_config.json
{
"mcpServers": {
"enterno": {
"command": "python",
"args": ["/path/to/enterno_mcp.py"]
}
}
} |
| TypeScript MCP server | import { Server } from '@modelcontextprotocol/sdk/server/index.js';
const server = new Server({ name: 'my-mcp', version: '1.0.0' }, { capabilities: { tools: {} } });
server.setRequestHandler(ListToolsRequestSchema, async () => ({
tools: [{ name: 'echo', description: 'Echo', inputSchema: { type: 'object', properties: { text: { type: 'string' } } } }]
}));
server.setRequestHandler(CallToolRequestSchema, async (req) => ({ content: [{ type: 'text', text: req.params.arguments.text }] })); |
| Claude Agent SDK (Python) | from anthropic import Anthropic
client = Anthropic()
# Claude Opus 4.7 with MCP tools support
response = client.messages.create(
model='claude-opus-4-7',
mcp_servers=[{'url': 'stdio://path/to/mcp.py'}],
messages=[{'role':'user','content':'Check DNS for google.com'}]
) |
| MCP over HTTP (remote) | # Start SSE-based MCP server
mcp = FastMCP('Enterno Remote', host='0.0.0.0', port=8000)
# Connect client via https://mcp.example.com/sse |
Типичные ошибки
- Tool descriptions vague — LLM не знает когда вызвать. Пишите precise: "Get weather for a specific city, returns JSON"
- Input schema пропустил required fields — LLM пропустит args. JSON Schema required обязателен
- No error handling — LLM получает exception string, confused. Return structured error: {error: "msg"}
- Long-running tools (>10s) — LLM timeout. Async с progress updates
- Not sandboxing — MCP server с shell access = RCE если prompt injection
Больше по теме
Часто задаваемые вопросы
MCP vs OpenAI function calling?
MCP: standard, works с multiple clients (Claude, Zed, others). OpenAI functions: specific к OpenAI API. MCP — более универсально.
MCP servers market 2026?
Официальная коллекция: filesystem, git, GitHub, Slack, Postgres, Google Drive. Community — десятки больше (Figma, Linear, Sentry).
Security?
Claude Desktop prompts user перед каждым tool call. Для automated agents — нужен manual review пайп, sandbox, whitelist tools.
Deploy MCP remotely?
SSE transport над HTTPS. Auth через OAuth. Пример — Sentry MCP server (opens at https://mcp.sentry.dev).