An AI that controls your computer, also called a computer use agent, is an AI that doesn’t just answer questions but does the work itself on your computer: it opens apps and websites, clicks buttons, types, fills in forms, moves files and sends messages. You ask in words or out loud, and it carries the task out step by step, like a person sitting at the screen.
The agent needs two things: to see what’s on the screen, and to act on it. There are two ways to see. The common one is screenshots: the model gets an image and guesses where to click by coordinates. The other is the operating system’s accessibility tree, the same information screen readers use. There every button and field has a name and a role, so the agent clicks “Send” by its name rather than a point in a picture. ands works the second way, and uses a screenshot only when needed and only with your OK.
How it differs from a chatbot
A chatbot writes you an answer, and you do the work. An AI that controls your computer does the work itself: instead of explaining how to fill in a form, it fills it in. Instead of drafting an email for you to paste, it opens your mail and gets it ready to send.
Cloud agent or an agent on your own computer
Some agents run on a remote computer in the cloud. Sites see a sign-in from another country and sometimes block it, the agent asks for your passwords, and you only see the result at the end. An agent that runs on your own computer, like ands, signs in from your network with the accounts you’re already signed in to, works in the apps you have installed, and lets you watch every step and stop it.
Why it matters where there’s no API
Most real work happens in software and sites that have no AI connection: old bookkeeping programs, government sites, university portals, CAD and design apps on perpetual licences. An AI that controls the computer works in them like a person, with the mouse and keyboard, with no plugin and no development.
How to use one safely
An AI that can click anything needs a brake. In ands every action is checked before it happens, by code outside the model: reading goes through, changes are covered by the plan you approved, and sending, paying, deleting, installing or signing in always wait for your click. Password fields are never read and password managers are blocked by default.
Things people ask
“Email my accountant the October invoices from the folder”
“Fill in the refund form on the site and stop before sending”
“Download every assignment from Moodle into the course folder”
“Export the frame as PNG at 2x”
“Tidy my desktop into folders by file type”
Questions
Yes. Apps like ands get Accessibility permission from the operating system, which lets them see buttons and fields and click and type. You grant it once and can take it back any time in your computer’s settings.
It is when there’s a check the model can’t get around. In ands anything sensitive, like sending, paying, deleting or installing, waits for your OK, and you watch every step on screen and can stop it.
No. The ands app is for everyone: download it, allow access and ask in English or Hebrew, by text or voice. Developers can connect ands as an MCP server to Claude Code, Codex or Cursor.
Yes. You can ask in Hebrew or English, by text or voice, and ands works on Hebrew sites and apps too.
The app is $20 a month with everything included and no other AI accounts. The MCP server and CLI for developers are free.
Macs with macOS 13 or later, Windows 10 and 11, and Linux.