Letting AI use tools
In one line: give a model the ability to search, read files or run commands and it stops being a thing that answers and becomes a thing that does — which is a much bigger step than it sounds.
What changes
A plain model can only produce text. Give it tools — search the web, read a file, run a command, call an API — and it can go and find out, then act on what it found.
That fixes the two biggest limits in one move. It no longer has to guess about anything recent, because it can look. And it no longer has to hand you instructions, because it can carry them out.
The word for one of these is an agent: a model in a loop, deciding what to do next, doing it, looking at the result, and going again until the job is done or it gets stuck.
What this looks like in practice
Search. Most chat tools can now look things up. It makes recency questions answerable and it usually cites what it read, which you should open.
Code assistants. Claude Code, Cursor, GitHub Copilot and friends read your actual project, change files, run the tests and read the failures. This is where the current tools are strongest by a distance.
Computer use. Newer, rougher: the model drives a browser or a desktop. Impressive and not yet reliable.
MCP (Model Context Protocol) is worth knowing as a name — an emerging standard for plugging tools into models, so a tool written once works with several of them. → modelcontextprotocol.io
The habit that keeps this safe
Let it propose, and you approve — at least until you trust the specific thing.
A model that is wrong in a chat costs you a paragraph. A model that is wrong with a file system costs you files. The good tools ask before they do anything irreversible, and that prompt is the feature, not the friction. Skipping it is the one shortcut worth not taking.
Two more:
Work where you can undo. A git repository, a copy, a virtual machine. Not because it will go wrong often, but because the recovery should be cheap when it does.
Read what it did. Not every line, but the shape of it. An agent that has been running unwatched for twenty minutes has made a lot of decisions you have not seen.
Where AIOS sits
AIOS is deliberately built this way round.
It acts on your notes deterministically, with no model involved. Capturing an event, finishing a task, moving something, rescheduling — those are rules, so they are the same every time and cannot be hallucinated.
It never deletes. AI can move things and mark them done; deleting is yours alone.
Everything it runs is logged. The Activity tab lists every command AIOS has run on this machine, with the reason, and that log is written before the command runs rather than after.
Claude Code is available inside it — the Assistant tab — and nothing it proposes reaches your notes until you press Keep.
Try this
Open the Activity tab and read the list. Every line is something AIOS ran on this PC and why. That is what "an AI tool you can audit" actually looks like, and it is worth expecting from anything you let near your machine.
If you want to go further
---
Next: Where to go next · All lessons