Model Context Protocol — открытый стандарт, который Anthropic опубликовал в ноябре 2024 для унификации discovery и invocation внешних capabilities AI-клиентов. Суть: один Server, подключение из Claude Desktop, Cursor или любого MCP-клиента — смена LLM без переписывания tool layer.

┌────────────────────┐         ┌─────────────────────┐
│   MCP Client       │ ◄─────► │   MCP Server        │
│  (Claude / Cursor) │  JSON   │  (вы строите это)   │
│                    │  -RPC   │                     │
└────────────────────┘         └─────────────────────┘
                                        │
                          ┌─────────────┼─────────────┐
                          ▼             ▼             ▼
                       Tools       Resources      Prompts
                    (действия)    (чтение data)  (шаблоны)

Wire format — JSON-RPC 2.0. Lifecycle: initialize handshake → capability negotiation → request/response loop → graceful shutdown. Два transport modes: stdio (локальный subprocess, zero network deps) и HTTP + SSE (remote, multi-client, team-shared).

Рабочий MCP Server за десять минут. Python с FastMCP — recommended starting path; TypeScript зеркалит концепции с @modelcontextprotocol/sdk.

python -m venv .venv
source .venv/bin/activate
pip install mcp

npm init -y
npm install @modelcontextprotocol/sdk

Рекомендуемый layout проекта:

my-mcp-server/
├── server.py
├── tools/
│   ├── calculator.py
│   └── web_search.py
├── resources/
│   └── file_reader.py
├── prompts/
│   └── templates.py
├── tests/
│   └── test_tools.py
└── pyproject.toml

from mcp.server.fastmcp import FastMCP

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

@mcp.tool()
def say_hello(name: str) -> str:
    return f"Hello, {name}! This is your first MCP tool."

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

Запуск и проверка через официальный MCP Inspector:

python server.py
npx @modelcontextprotocol/inspector python server.py

Подключение клиентов: launch command в Claude Desktop claude_desktop_config.json или Cursor mcp.json. После restart — tool в списке и test invocation.

Production Server регистрирует Tools для actions, Resources для readable data и Prompts для reusable templates. Function signatures и docstrings — AI-facing documentation без отдельного OpenAPI.

Tool structure: name из function, parameters из type hints, description из docstring. Pydantic models для complex inputs:

from pydantic import BaseModel, Field

class SearchInput(BaseModel):
    query: str = Field(description="Search keywords")
    max_results: int = Field(default=5, description="Max results")
    language: str = Field(default="en", description="Result language")

@mcp.tool()
def web_search(input: SearchInput) -> list[dict]:
    ...

import httpx

@mcp.tool()
async def fetch_url(url: str) -> str:
    async with httpx.AsyncClient(timeout=30.0) as client:
        response = await client.get(url)
        response.raise_for_status()
        return response.text

Error handling: structured error dicts для recoverable failures; raise только для unrecoverable. Timeouts на I/O tools. Validate permissions перед filesystem или database access.

Resources отличаются от Tools: Resources supply data via URI, Tools execute actions. Static resources — fixed URIs; dynamic — path parameters.

@mcp.resource("config://app-settings")
def get_app_settings() -> str:
    return json.dumps({"version": "1.0", "env": "production"})

@mcp.resource("user://{user_id}/profile")
def get_user_profile(user_id: str) -> str:
    user = db.query_user(user_id)
    return json.dumps(user)

Filesystem resource Server lists directories, reads files, subscribes to change events. Prompts ship pre-built conversation templates с dynamic parameter injection:

from mcp.types import PromptMessage, TextContent

@mcp.prompt()
def code_review_prompt(language: str, code: str) -> list[PromptMessage]:
    return [
        PromptMessage(
            role="user",
            content=TextContent(
                type="text",
                text=f"Review this {language} code for bugs, security, and performance:\n```{language}\n{code}\n```"
            )
        )
    ]

Multi-turn Prompts mix user и assistant roles — interview simulations, debug walkthroughs, onboarding flows с consistent structure.

Local STDIO для dev; team-shared или SaaS deployments нужен HTTP + SSE. FastMCP supports streamable-http с minimal code change.

mcp = FastMCP("remote-server", transport="streamable-http")

if __name__ == "__main__":
    mcp.run(host="0.0.0.0", port=8000)

Production HTTP Servers нужны auth и guardrails: Bearer Token или API Key middleware, CORS restricted to known origins, rate limiting на tool endpoints с external APIs или databases.

MCP Inspector — primary debug UI: launch against Server command, call Tools interactively, inspect JSON-RPC pairs, simulate error paths перед wiring client.

import pytest
from mcp.client.session import ClientSession
from mcp.client.stdio import StdioServerParameters, stdio_client

@pytest.mark.asyncio
async def test_calculator_tool():
    server_params = StdioServerParameters(
        command="python",
        args=["server.py"]
    )
    async with stdio_client(server_params) as (read, write):
        async with ClientSession(read, write) as session:
            await session.initialize()
            result = await session.call_tool("calculate", {"expression": "2 + 2"})
            assert result.content[0].text == "4"

От Hello World к production: containerization, observability, version negotiation и reference project с real value. Capstone — personal knowledge-base MCP Server: semantic search over local Markdown notes в Cursor.

Deployment options: Docker wraps Python deps. Railway и Render — one-click deploys. AWS Lambda и Google Cloud Run для bursty serverless. Self-hosted VPS + Nginx для teams с existing infra. Structured logging, Prometheus metrics на tool calls, Sentry alerts, /health endpoint. Declare MCP protocol version, negotiate capabilities at handshake.

Knowledge-base project: Index local Markdown с ChromaDB или Qdrant, embed с text-embedding-3-small, watch changes via watchfiles. Три Tools — index notes, semantic search, write/update note — плюс Resource listing indexed files. В Cursor: «Что я писал про MCP на прошлой неделе?» — Agent calls search tool.

В 2026 все major AI clients ship native MCP support. Marketplace directories — thousands of Servers; enterprise adoption drives OAuth 2.0 и security audits. Next steps: spec на modelcontextprotocol.io, publish first Server, combine MCP с Agent Skills.

Сравните alternatives честно. HTTP+SSE на primary MacBook breaks при lid close и OS updates. Serverless без session affinity drops multi-step Agent context. Exposed MCP endpoints без auth — prompt-injection и unauthorized tool calls; ~1 000 unprotected Servers в 2026 audits. Для teams с Apple Silicon, 24/7 uptime и isolation между MCP Server, iOS CI и Agent Gateway workloads аренда dedicated KVMNODE Mac Mini M4 или M4 Pro — often better answer: flexible daily/weekly/monthly terms, launchd-managed processes, snapshot rollback. Tiers на странице цен, setup в центре помощи, provision через заказ когда MCP stack needs host that stays up after laptop close.