What is an MCP (Model Context Protocol)?

What is an MCP (Model Context Protocol)?

The MCP, for Model Context Protocol, is an open standard proposed by Anthropic in late 2024. Its role: to link an AI model to external tools, data and services in a uniform way. In other words, it gives the AI a standard way to step out of its window and reach the real world.

The most common image is the USB-C of AI. Before USB-C, every device had its own cable; with it, one port is enough. The MCP plays the same role: instead of a custom-coded integration for each tool, it offers a universal port that all compatible tools can use.

Be careful not to confuse two words. The MCP protocol is the standard, the rules of the game. An MCP server is the concrete connector that applies that standard for a specific tool, for example an MCP server for GitHub or for a database. The protocol is the language, the server is the counterpart.

How does an MCP server work?

The MCP relies on a simple client-server architecture. Your assistant (Claude, Claude Code, Cursor) acts as the client. The MCP server is the connector that exposes three things to the AI: tools (actions to perform), resources (data to read) and sometimes ready-made prompts.

A request's journey is concrete. The AI discovers the tools the server makes available, picks the right one, calls it with parameters, receives a structured result, then acts. You simply describe the need in plain language.

The key word is standardised. Every MCP server speaks the same language, so the AI does not have to learn a different integration for each tool. That is what makes the system both easy to extend and interoperable across AI coding tools.

What is an MCP used for day to day?

In practice, an MCP server lets the AI act where your work happens. It can read and post in Slack, open and comment on issues or pull requests, read and update Notion pages, or query a database in plain language, without you leaving your assistant.

The point is not the gadget, it is access to real context: the AI works on your actual data instead of guessing. Here are three examples of MCP servers, from code to issue tracking and databases, to show the variety of what you can plug in:

What is the difference between an MCP an API a skill and a plugin?

These words look alike but do not play the same role. An API is the raw pipe of a service: a developer has to integrate it by hand, tool by tool. The MCP is the standard layer that makes those APIs consumable by any AI, without rewriting the integration each time.

A skill is a how-to that frames a task (audit a page, debug, document). A plugin bundles several skills and commands. The MCP, on the other hand, is a connection channel to the outside world: it gives the AI access to tools, where the skill and the plugin frame its method.

In short: the MCP brings the access, the skill and the plugin bring the know-how. The three complement each other rather than compete. An ideal assistant is plugged into your tools by MCP servers and equipped with well-chosen skills.

Why does the MCP change the game?

Before the MCP, each vendor coded its own integrations, not reusable from one tool to another. It was the 'everyone has their own cable' problem: as many integrations to write and maintain as there were tool-assistant pairs.

With a common open protocol, an MCP server written once works in Claude, Claude Code, Cursor and any compatible client. An ecosystem of shared connectors forms and grows fast, cuts integration work and avoids lock-in to a single vendor.

It is the same leap as USB or HTTP in their day: a quiet standard that unlocks a whole market. That is why the MCP, launched in late 2024, was adopted in a short time by the main players in AI.

What precautions should you take with an MCP server?

An MCP server receives real access rights to your data and your tools. So caution is in order: check its source before installing it, start read-only, limit permissions to the strict minimum and keep an eye on what it can do.

The standard is also young: launched in late 2024, the ecosystem moves fast and not all servers are equal. Some are official and well maintained, others are not. And as always, the AI can misuse a tool it has access to.

The rule stays the same as for the rest: human review and scoping access are not optional. A well-chosen, well-framed MCP is a powerful lever; opened without judgement, it is a risk.

How do you use your first MCP?

Start small: pick one tool you check every day, plug it in read-only first to see what the AI gets from it, then open write access once the scope is validated. Test on a real case before multiplying connections.

Choosing the specific connectors to install, family by family, is covered in a dedicated guide. To go further, explore the connectors and integrations category of the catalogue and plug in your first MCP server.