What Is MCP? The Model Context Protocol, Explained in Plain English

MCP (Model Context Protocol) is an open standard that lets AI models use external tools and data — your databases, your APIs, your file system — through one common interface. Think of it as USB-C for AI applications: one connector, everything plugs in.

The problem MCP solves

Large language models are smart but isolated. Out of the box, a model can't read your database, check your calendar, or call your internal API. Before MCP, every AI application solved this with its own custom plugin system — which meant if you built a tool integration for one assistant, you had to rebuild it for every other one.

That's an N×M problem: N applications times M tools equals N×M custom integrations. MCP collapses it to N+M. You build your tool as an MCP server once, and every MCP-compatible application — Claude Desktop, Claude Code, VS Code, Cursor, ChatGPT, and more — can use it.

How MCP works

MCP has three participants:

• Host / client — the AI application the user interacts with (e.g., Claude Desktop or VS Code). It manages connections to servers and decides when the model should use a tool.
• MCP server — a small program you write (or install) that exposes capabilities. It might wrap a database, a SaaS API, or local files.
• The model — sees the tools each server offers, and asks the client to call them when needed.

What a server can expose

• Tools — functions the model can call, like search_orders(customer_id) or create_ticket(title, body). This is the workhorse of MCP.
• Resources — data the client can read into context, like file contents or database schemas.
• Prompts — reusable prompt templates the server offers to users (these power slash commands in some clients).

How they talk

Under the hood, clients and servers exchange JSON-RPC 2.0 messages over one of two transports:

• stdio — the client launches the server as a local subprocess and talks over stdin/stdout. Simple, no network, great for local tools.
• Streamable HTTP — the server runs as a web service. This is how remote, shared MCP servers work, typically secured with OAuth.

When a client connects, the two sides negotiate capabilities, the client asks the server what tools it has (tools/list), and from then on the model can request calls (tools/call) as it works. You rarely touch this layer directly — the official SDKs handle it.

A concrete example

Say you ask Claude: "Which of our customers churned last month?" Without MCP, Claude can only guess. With a Postgres MCP server connected:

1. Claude sees the server exposes a query tool.
2. It writes the SQL and asks the client to run query("SELECT ...").
3. The MCP server executes it against your database and returns the rows.
4. Claude reads the result and answers with real data.

The same pattern works for GitHub, Slack, Stripe, your internal admin API — anything you can wrap in a server.

Why MCP won

MCP was released by Anthropic in November 2024 and was adopted unusually fast — OpenAI, Google, Microsoft, and most major AI dev tools added support within a year. The reasons are practical:

• It's open — no vendor lock-in, open-source SDKs in TypeScript, Python, Java, C#, and more.
• Write once, run everywhere — one server serves every compatible client.
• It composes — a client can connect to many servers at once, and the model orchestrates across all of them.

What to learn next

The best way to make MCP click is to build a server — it takes about 30 minutes: Build Your First MCP Server in TypeScript. If you're wondering how MCP relates to the tool calling you may already use in LLM APIs, read MCP vs Function Calling.

FAQ

What does MCP stand for?

Model Context Protocol — an open standard, originally created by Anthropic, for connecting AI applications to external tools and data.

Is MCP free to use?

Yes. The protocol is open, and the official SDKs are open source. You can build and run servers for free.

Client vs server — which is which?

The client is the AI app (Claude Desktop, VS Code, Cursor). The server is the program exposing tools. One client can connect to many servers.

Which AI apps support MCP?

Claude Desktop, Claude Code, VS Code (GitHub Copilot), Cursor, Windsurf, ChatGPT, and a growing list of others.