Model Context Protocol

A chatbot only knows by itself what was in its training data. It knows neither your files nor today’s calendar nor a company’s database. For it to access these, it needs a connection to the outside world. The Model Context Protocol is a fixed set of rules for exactly such connections. It describes how an AI program asks for available data and tools and in what form it gets answers back. It was developed in 2024 by the US company Anthropic and released as an open standard that anyone is free to use.

Why this matters

Without a shared standard, things quickly turn into a mess. Every AI application would need its own integration for every third-party program. With ten applications and ten programs, that would be a hundred individual solutions. With a shared protocol, it’s twenty: each side follows the rules once.

This is the same idea as with USB. Before USB, every device had its own connector. Afterward, one cable fit everywhere. For companies, this is a cost factor. They don’t want to have to rebuild their AI tools whenever they switch providers.

The standard is also important for agents. This is the term for AI systems that don’t just respond but carry out steps themselves, such as saving a file or checking an order. Such systems are only useful if they can reliably reach real tools.

How it works

The protocol divides the world into two roles. On one side is the client, meaning the program in which the AI runs. On the other side is the server, a small auxiliary program that provides a specific data source. There are servers for file folders, for calendars, for databases, or for ticketing systems.

At startup, the client contacts the server and asks: What can you do? The server responds with a list. It states, for example, that it can read and search files. Each entry comes with a description in plain language. The AI model reads these descriptions and uses them to decide which tool fits a given question.

After that, the exchange runs in clearly formed messages. The AI requests an action, the server carries it out and sends the result back. This result ends up in the model’s context, meaning the text it takes into account when formulating its answer. That is where the protocol’s name comes from.

One point remains delicate: whoever connects a server grants the AI real permissions. That’s why many programs ask for confirmation before every access. Otherwise, a manipulated server could siphon off data or feed the model false instructions.

Where you encounter the term

MCP is most commonly encountered in programming tools. Editors like Cursor or VS Code use it to connect the project folder, version control, or documentation to the AI assistant. Desktop apps from major providers now support the protocol as well.

The term appears in business news when it comes to standards and market power. In 2025, several major providers announced support for MCP, including OpenAI and Google. Such announcements are relevant because whoever’s standard prevails gains an advantage. Users who only operate chat windows, however, never see MCP directly. It works in the background, similar to the protocols through which your browser loads web pages.