Computer use

Computer use lets the assistant open a real web browser and work a website for you: navigating, clicking, typing, reading the page, scrolling, filling in fields and waiting for things to load. You watch it happen live in a panel beside the chat, and you can take the controls at any moment.

This page covers what the assistant will use a browser for, how a task starts, what you see while it runs, what it will not do on its own, and the shape of what it costs. For stepping in mid-task and for signing in to a site, see Taking control and signing in.

What it is for

Most questions do not need a browser. If something can be answered by reading a page, the assistant reads it — that is faster and cheaper, and it tries that first. It deliberately stays out of the browser for anything a plain read answers: a fact, a definition, today's news, a summary of a page it can simply fetch.

A browser is for the jobs where reading is not enough and someone has to actually work the site: click through it, sign in, filter, sort, page through results, fill something in. The assistant reaches for one when a request looks like:

• Book, order, reserve, apply, submit or register for something.
• Check an account, a balance, an order, a booking, a tracking number or an invoice.
• Find flights, hotels, tickets, appointments, listings or prices to compare.
• Fill in a form.
• Look through a portal, dashboard or inbox that has no other way in.
• Download something that sits behind a login.
• Anything phrased as “can you sort out X for me”, where X lives on a website.

How a task starts

You do not have to ask for a browser by name. The assistant decides for itself, using a simple test: could a person get this done by reading a page, or would they have to work the site? If it is the second, it opens a browser and gets going.

There is no approval step and no card to confirm for an ordinary browser task. The panel opens by itself and the work starts. Signing in to a site is the one exception — that always asks you first.

The assistant will also fall back to a browser when an ordinary lookup comes back thin, out of date or blocked, or when a page will not give up its content without a real browser. So “that answer was useless, go and actually look” works as a follow-up.

Keeping it to particular sites

You can scope a task to named sites — “only look on these two”. That is a real restriction rather than a suggestion. If a link, a redirect, a new tab or a tab switch would take it somewhere else, it stops instead, names the sites it is allowed on and the one it stopped at, and invites you to add it.

What you see while it works

A panel slides in from the right of the chat, headed Computer. It opens on its own the moment a browser task starts. Inside it is the browser itself, live and moving, and it refreshes continuously — there is nothing to reload.

A status chip in the panel header tracks the run: Starting, then Working, and later Done, Stopped, Failed, Asleep, or Yours once a run has ended and the browser is sitting there waiting for you.

While it is actually working, the chip stops saying Working and narrates the current step instead, in plain words: Opening and the address, Reading the page, Looking for and what it is after, Clicking, Typing, Filling in a field, Pressing, Scrolling, Waiting, Opening a tab, Switching tab. You can read what it is doing, not just that it is busy.

The card in the chat

A card also appears in the conversation for that run. Its headline follows the state — Opening the browser, Working in the browser, Done — browser still open, Couldn't finish, Stopped, Browser closed, Asleep — tap to wake — and while the run is going, the headline carries the same live narration, with the state word moving to the line underneath. That second line also shows how many actions it has taken so far and the address the browser is on.

The card's button changes with the state: Watch while it is working, Open while you are driving, Wake when the browser is asleep, and Reopen once the run has ended.

On different screen sizes

On a wide screen the panel sits beside the conversation, and an expand control makes it wider still; the app's side navigation tucks away to give it room and comes back when you shrink it again. Your choice is remembered on that device. On a narrower screen the panel slides over the conversation instead, tapping outside it closes it, and the expand control is not shown.

The message box waits

While a browser task is running, the conversation waits for it. The send button is replaced by a stop button and the message box reads Working in the browser…; sending anyway gets you Still working in the browser — one moment.

What happens instead is that the assistant replies straight away to say what it has set going, the task runs on its own, and it comes back with the outcome as a new message when it is done. Ask for the next thing then — it will carry straight on in the same browser.

One browser that stays yours

The browser belongs to you, not to the task. A follow-up request continues in the same browser, with the same tabs, the same history and the same sign-ins. It is tied to your account rather than to one conversation, so opening a new chat still picks up the same computer.

When a task finishes, the browser does not close. It is deliberately left open on the page it reached so you can carry on from there by hand or ask for the next thing.

After a stretch with nobody using it, the browser puts itself to sleep. Asleep is not closed: your sign-ins, cookies, site preferences — down to a cookie banner you have already dismissed — and the page it was on are all kept. The panel shows Your computer is asleep. with a line telling you where it will wake up.

Waking is automatic. Ask for the next thing and the assistant wakes it itself: it wakes up where you left it, still signed in, with nothing for you to do first. You can also press Wake it up in the panel or Wake on the card.

What it will not do on its own

Some things are off limits to the assistant however the request is worded. They are enforced underneath it, so rephrasing or taking a different route to the same place does not get past them. When it hits one, it says plainly what it stopped at and asks you to do that part yourself in the panel.

• It will not type a password, a passcode, or a security, one-time or two-step code. It stops at the field and asks you to enter it yourself in the live view so it can carry on.
• It will not enter payment card numbers.
• If it cannot tell what a field is for, it will not type into it at all rather than guess.
• It never attempts a CAPTCHA or a “confirm you're human” check. The run parks and waits for a person, the browser stays open on that page, and you clear the check and tell the assistant to continue.
• It stops before an action that would commit you to something — spending money, placing an order, sending something out, or making a change that would be hard to undo. Unless your own request asked for that specific thing, it tells you what it is about to do and waits for you to confirm. That stop is per action, so it will fill a form in and then stop short of actually submitting it.
• It only opens ordinary web pages, and it works a site with its normal controls only — the same clicks, keystrokes and scrolling you would use.
• It does not take instructions from web pages. If a page contains text written to look like a request to the assistant, it is treated as page content rather than something to act on, and the assistant tells you it saw it.

None of these stops is a failure, and none of them closes the browser. It stays open on the page it reached so you can finish that step yourself and tell the assistant to carry on. If a run ends after one of them, the assistant's closing message says what it was stopped by rather than reporting a half-finished job as done.

The assistant only, not agents

Computer use belongs to the assistant in chat. Agents — the scheduled, unattended automations that run without you there — cannot drive a browser at all. They also cannot see which sites you have saved sign-ins for, and cannot start a sign-in.

That is deliberate. A real browser session runs into things that need a person: a check to clear, a site asking you to sign in again, a page that has changed shape since the task was written. An unattended run has nobody to do any of that, so rather than half-support it, agents get no browser. An agent asked for browser work says so in its result and hands the job back to you instead of failing quietly, and it cannot pick up where your browser left off.

Computer use is also never reachable by someone else talking to your assistant from the outside, such as an inbound caller or an inbound email conversation. The browser always runs as the account owner — and, with a saved sign-in, as the account owner already signed in.

What it costs

A browser task uses credits while it works. The time the browser is open and the assistant's work of driving it are billed together as one charge for that browser session, not per action.

The thing worth knowing is that the meter runs for as long as the browser is open, not only while the assistant is working. That includes the stretch after a task has finished while the browser sits waiting for your next request, and any time you are driving it yourself.

It stops when the browser actually closes: when you press Stop, or when the browser puts itself to sleep after a short stretch with nobody using it. Asleep costs nothing and loses nothing. Waking is not free, though — waking starts a browser again, and the meter with it.

Having the panel open does not by itself hold the browser awake. It only stays awake while you are actually using it, so a panel left open on a screen you are not looking at will not keep it running.

On your usage breakdown this appears as its own line, labelled Computer use.

How a run ends

When the task is done, the assistant comes back into the conversation with the outcome as a new message, on its own — you do not have to ask how it went — and the browser is left open. A run can also end in these ways:

• You pressed Stop. That is recorded as you stopping it, not as a failure.
• It handed over to you — a login wall, a “confirm you're human” check, a step that would change something, or you taking the controls. The browser stays open and this is not a failure either.
• It ran out of time. There is a limit on how long a single task can run, and a long job can reach it before it is finished.
• It reached the spending limit for that run. The browser is left on the page it reached and the assistant tells you what still needs doing.

If computer use is not available on your account, the assistant tells you so when you ask for it.

Next steps