MCP explained simply: connect AI to the tools you use
What the Model Context Protocol is, in plain language, and the three connections worth making before any of the clever ones.
The problem it solves
A model can only work with what you paste in. That is fine for one document and hopeless for a calendar, a file server or a ticket system you touch fifty times a day.
MCP is a common way for a model to ask a tool for something and get a structured answer back. Instead of one custom integration per app, tools speak one protocol and any client can use them.
What a server actually is
The word server sounds heavier than the thing. In practice it is a small program that exposes a few named actions.
Three connections worth making first
Start where you lose the most time to copy and paste, not where the demo looks best.
What to be careful about
Every connection you open is a door. Grant read before write, keep write access to systems where a mistake is reversible, and never connect production billing to an experiment.
Ask for a plan before an action whenever the action is not trivially undoable. A confirmation step costs two seconds and saves an afternoon.
Before using any tool that writes, changes or sends: summarise what you are about to do in one line and wait for my confirmation. Read only lookups do not need confirmation.
When you do not need it
If your work is one document at a time, MCP adds setup and gives you nothing. Paste the file and move on.
The moment you notice you are the integration, moving the same data between two apps by hand every day, that is when the setup pays for itself.