Español

Lee este artículo en español.

What is MCP, and why your AI assistant can now use real tools

Ask an AI assistant to "make me a QR code" or "turn this into a PDF," and until recently the honest answer was: it can tell you how, but it can't actually do it. It has no hands. It can write code for you to run, or describe the steps, but it can't reach outside the conversation and hand you back a real file.

MCP (Model Context Protocol) is what changes that. It's an open standard, published by Anthropic, that lets an AI assistant connect to outside tools and actually use them mid-conversation, the same way your phone connects to a printer without you having to explain what a printer is every time.

In plain terms: a menu the assistant can read

An MCP server is a small list of things it can do, written so the assistant can understand it: "generate a QR code, you give me the text, I give you back an image" is one entry on that list. When you ask for a QR code, the assistant doesn't guess or write code: it reads that entry, calls it with what you asked for, and gets back a real result, which it then shows you or hands you a link to.

The assistant still does the thinking (understanding what you want, choosing which tool fits). The MCP server does the doing: the actual PDF gets generated, the actual screenshot gets taken, the actual QR code gets built, by code written for exactly that job, not by the AI guessing at pixels.

A live example, not a hypothetical one

Cofferdock runs its own MCP server, and it's the one this article's examples are drawn from, not a made-up scenario. Once connected, you can ask Claude (or another assistant that supports MCP) things like:

and get back an actual file, inside the same conversation, drawing from the same 25 tools available through the regular API: PDFs, images, QR codes, screenshots, charts and a handful of validators.

Connecting without handing over a password

The simplest way in is "Sign in with Cofferdock": you click connect, log in the way you'd log into any website, and allow access, the same flow as signing into an app with your Google account. Behind it, the assistant never sees your password: it gets a temporary key tied to your account, which renews itself on its own and which you can revoke at any moment from your dashboard, without changing anything else. Each call still spends a credit from your account, exactly like calling the API directly.

FAQ

Do I need to know how to code to use this? No. Connecting is a login and a click. Once it's connected, you just ask for what you need in plain language, the same as any other conversation.

Is this the same as giving the AI my API key? Not with "Sign in with Cofferdock": no key ever changes hands, you log in directly and it's revocable any time. An API key is still an option too, mainly for tools that don't support the login flow yet.

Which assistants support this today? Claude (web, mobile and desktop), Claude Code, Cursor, VS Code with Copilot, and n8n (as a tool inside an AI Agent node) all support connecting to an MCP server today. Support for more assistants is expanding as MCP itself grows.

Try it

The full setup, with the exact steps for each assistant, is on the AI assistants page. If you don't have an account yet, you can create one for free.