Connect Claude to
live information

Back in Chapter 11 you learned the model\u2019s knowledge is frozen at a past date, so anything current is a guess. This chapter is the fix. You can plug the model into live sources so that when accuracy matters, it reads the real, current thing instead of reciting its stale memory. The plug has a name: MCP.

MCP stands for a connection standard, but you do not need the words. People will say "MCP" at you like it is a password to a club you are not in. It is not. You need the picture, not the acronym. Think of it as a USB-C port for AI. Before USB-C, every device had its own awkward cable, and you owned a drawer full of them, all slightly wrong. Then one standard port meant anything could plug into anything. MCP is that for AI: one standard way to connect the model to outside tools and live data, so it can reach beyond its frozen memory.

A few more of those club words, since this is where they start showing up, and none of them are as technical as they sound. An API is just a way for one program to talk to another, which is exactly what a connection like this uses under the hood. A key is a password that proves a request is really yours, the thing you paste in once so a tool knows to let you through. You do not need to touch either to use a Connector here, but you will hear the words, and now they are demystified instead of intimidating.

Why this matters is simple and it ties the whole judgment section together. The model is brilliant but frozen. Its training stopped on a certain day, and the world kept moving. Ask it about a current price, a new version, today\u2019s documentation, and it will answer confidently from memory that may be a year stale. A live connection changes the answer to when accuracy matters, give it a source, not its memory.

You do not have to build anything to use one. Connecting a source is a setup step you do once, and then the model can reach it whenever a task needs it. The shape is always the same.

Here is the one a beginner can actually try today. It is called context7, and it pulls current documentation for software tools, the kind of thing the model is most likely to be out of date about. You add it once, then you literally just write "use context7" in a request and it fetches the live docs.

That last line is the real lesson hiding in a setup chapter. Even for things the model thinks it knows, the honest move is to check the current source, because training data goes stale quietly. A live connection is not only for things the model has never heard of. It is for everything where being current actually matters, including the things it would happily answer wrong from memory.

context7 is the gentle first try, but be honest: most people do not spend their day reading software documentation. The connection you will reach for again and again is the one that plugs the model into the place your real work already lives. For most people that is Google Drive.

In Claude this lives in a panel called Connectors. You will find it in settings, a tidy list of things you can switch on: Google Drive, calendars, and more arriving steadily. Toggle Google Drive on, sign in once the way you sign in to anything, and from then on you can say "look in my Drive" and the model reads your actual documents instead of asking you to paste them in one by one like it is 2009.

That is the whole shift. A Connector turns "let me find that file, open it, copy the bit you need, paste it, hope I got the right version" into one sentence. The model reaches into the real source and works from the true, current thing. That last line in the prompt matters for the same reason it always does: you want the file, not a confident guess about the file.

The worry that stops people connecting anything is the right worry to have, and the answer is reassuring on both counts. First, reach: a connector can only get to what you can already get to. It borrows your own permissions from the tool you connect, it does not widen them, so if you cannot open a file, a channel, or someone else's inbox yourself, the connector cannot either. Claude sees what you see, and nothing past it. Second, the privacy line most people get backwards: a file Claude reads through a connector is not used to train the model. The only catch is the obvious one. If you copy that file's contents straight into the chat yourself, it is then treated like anything else you typed, and the ordinary training rule from Chapter 5 applies, the one you control with a setting. So let it read through the connector when you can, and you keep the cleaner privacy line.

Other tools, same idea. This is not a Claude-only trick. ChatGPT has connectors and a Drive link too, Gemini reaches into Google's own apps, and the MCP standard itself is shared across tools. The button is in a different place and the name on the panel changes. The judgment, plug the model into the real source instead of trusting its memory, travels with you to whichever tool you end up using.

The model knows an enormous amount, frozen on the day it stopped learning. The port that lets it reach today is one setup away, and knowing when to use it is the judgment that keeps you honest.