What is MCP (Model Context Protocol)?

The Model Context Protocol (MCP) is an open standard for connecting AI applications to the tools and data they need. Anthropic introduced it in November 2024 and published it as an open specification.

The usual analogy is a USB-C port: one connector, many devices, no adapter drawer.

The problem it solves

Before a standard existed, connecting AI applications to tools was a multiplication problem. Every application needed its own integration with every data source — build them all separately, maintain them all separately, and repeat the whole exercise for each new application.

A shared protocol turns multiplication into addition. A tool implements MCP once and any MCP-speaking application can use it. An application speaks MCP once and gets every MCP server ever written.

That's the entire pitch, and it's a good one — it is the same argument that made USB, HTTP and SQL win.

How it fits together

• MCP server — wraps a capability: your files, a database, a Git repo, an API.
• MCP client — the AI application: an assistant, an IDE, an agent.
• The protocol — how they talk. Servers advertise what they offer; clients call it.

Servers expose three kinds of thing: tools (actions the model can take), resources (data it can read), and prompts (reusable templates).

Why you might care

If you use AI tools, MCP is why an assistant can suddenly read your project files or query your database without anyone building a bespoke feature for it. The integrations arrive from an ecosystem rather than a roadmap.

If you build with AI, it means writing an integration once rather than once per host application.

What it does not do

MCP is plumbing, and plumbing solves plumbing problems:

• It doesn't make the model smarter. A model with database access still writes bad queries if it wrote bad queries before.
• It doesn't solve permissions. Connecting a model to a system means deciding what it may do there. The protocol carries the request; it doesn't decide whether the request is wise.
• It doesn't remove the risk of tool use. A model that can act can act wrongly. A standard connector makes that easier to arrange, not safer.
• It doesn't guarantee quality. Anyone can publish a server. Connecting an untrusted one to your assistant is a supply-chain decision, and content returned by a server is data, not instructions — a distinction worth keeping firmly in mind.

The protocol is genuinely useful and genuinely boring, which is the correct character for infrastructure.