MCP (Model Context Protocol)

The Model Context Protocol is an open standard that lets developers build secure, two-way connections between data sources and AI-powered tools. It was introduced by Anthropic on November 25, 2024, and is now broadly adopted across Claude Desktop, Claude Code, Visual Studio Code, Cursor, and many third-party IDEs. MCP follows a client-host-server architecture where each Host runs multiple Client instances, and each Client maintains a 1:1 stateful session with one Server. The protocol uses JSON-RPC 2.0 as its message format and supports two transports: stdio for local servers and Streamable HTTP for remote servers. The current spec version is 2025-11-25.

MCP defines three server-exposed primitives: Prompts (user-controlled templates), Resources (application-controlled data the model can read), and Tools (model-controlled executable functions). The lifecycle is initialize → operate → shutdown, with capability negotiation during initialization. The win is interop: write one MCP server for your internal APIs, then any MCP-compatible client across any LLM provider can use it without per-client integration code. Gartner predicts that by the end of 2026, 75% of API gateway vendors will include MCP support in their offerings.

Canonical MCP topology per the official spec. Hosts manage multiple Clients. Each Client has a 1:1 session with one Server, exchanging JSON-RPC 2.0 messages. Remote Servers can be shared across Clients.

// Minimal MCP server exposing one tool primitive
import { Server } from "@modelcontextprotocol/sdk/server/index.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";

const server = new Server({ name: "weather", version: "1.0.0" });
server.setRequestHandler({ method: "tools/list" }, async () => ({
  tools: [{
    name: "get_weather",
    description: "Get current weather for a city",
    inputSchema: { type: "object", properties: { city: { type: "string" } } },
  }],
}));
await server.connect(new StdioServerTransport());

Local MCP server using stdio transport. Any MCP-compatible Host (Claude Desktop, Cursor, VS Code) can connect.

When you want one tool/resource definition to work across multiple LLM clients, when exposing internal APIs to AI tools without per-client code, or when you need standardized agent-to-data connections that survive vendor changes.

• MCP official site
• Anthropic announcement (Nov 2024)
• Architecture spec
• Spec versioning

Related Terms