What gatto does

What a local AI model is

An AI model is a file. A big one — a few gigabytes — and it sits on your hard drive like any other file. Your graphics card reads it and produces the answers.

That is the whole trick. There is no server on the other end. Once the file is on your disk, nothing you type and nothing you open has to leave the machine: no account, no subscription, no request going anywhere. You can unplug the network and keep working.

The catch is that a model has to fit. A model too big for your graphics card will not run, or will run so slowly that you would not want it to. Which model your PC can actually handle is the first thing gatto works out for you, and it is where the rest of this page starts.

What setup does

The first time you run gatto it takes you through four things. You answer questions; it does the work.

1. It finds the engine your card can use. Running a model needs a small program underneath, and there is a different build of it for each kind of graphics card. gatto looks for one you already have and offers to fetch the matching build if you do not.
2. It finds a model that fits. It reads how much memory your graphics card has, does the arithmetic, and only shows you models that will actually run on your hardware. You pick from what fits rather than finding out afterwards.
3. It writes nothing until the last screen. Every answer you give is held. Close the window halfway through and there is nothing to clean up — no folders, no settings, no half-finished configuration. The wizard only commits when you reach the end and say so.
4. It shows you that it works. When everything is in place gatto asks your new setup a question and shows you the reply. You watch it answer before you are handed the prompt.

After that, adding a second model is one command, and switching between them is one more. You are not locked into the first choice you made.

What it can do for you

Six things come built in. In plain words, gatto can:

• read a file and tell you what is in it
• write a new file from scratch
• change part of a file without rewriting the rest
• find files by name, anywhere under the folder you opened
• search inside files for a word or a pattern
• run a command and read what it printed

That is a small list on purpose. Almost everything you would want done to a project is one of those six, and a smaller model does better work when it has fewer things to choose between.

More arrives depending on how you are set up: gatto can keep notes about a project between sessions, hand a long job to a second copy of itself so your own conversation stays short, and — through the extensions that ship with it — fetch a web page, search the web, or stop and ask you a multiple-choice question. Extensions are how it gains anything else; the extensions page covers them.

How it asks permission

Reading is free. Anything that changes something stops and asks first: writing a file, editing a file, running a command. You see exactly what it wants to do before it happens, and you can allow it once, allow that kind of thing from then on, or refuse.

The other half matters more and is easy to miss. gatto only loads extensions from your own home folder — never from the project you happen to have open. So opening someone else's code can never hand gatto new powers, no matter what that code says. What gatto is able to do is decided by you, on your machine, before you ever open the folder.

Checking a model can do the job

Small AI models are uneven in ways that are hard to see from the outside. A model that writes beautifully and cannot reliably use a tool is no use as an assistant, and you cannot tell which one you have by reading its description.

So gatto checks. It offers to try your new setup during setup, and you can run the same check later with gatto audition: a fixed set of tasks through gatto's real tools, marked mechanically. No second AI judging the first, no opinion, no score out of ten — it is a check that the thing works, and it either does or it does not.