Browser

When a task needs a website rather than an API, the model opens a browser of its own. It navigates, reads pages, including embedded chat widgets, clicks, types, fills forms, takes screenshots, and downloads files. Every step lands in the transcript as a compact row: the page as the model read it after each step sits behind the row, and screenshots open in the preview.

The Browser pane

The Browser button in the chat's top bar opens the live browser beside the transcript (full screen on a phone). It opens on its own the first time a turn uses the browser, and while the pane is closed the button's dot says what is happening: a steady dot for an open browser, a pulsing one while the agent is driving, a halo while a person is.

Watching and taking over

While the agent drives, you watch: the frame glows and ignores clicks so nobody fights the agent. Take over hands you the wheel, at once when the browser is idle or, during a turn, as soon as the current step finishes, and the turn waits for you. Hand back returns control, and so does sending a message, since the turn that starts is the agent's. Teammates in the chat see who is driving.

When the browser needs you

Sign-in walls, verification codes, and anything the model should not decide alone come to you: the chat shows a card and your Inbox gets a row. Sign in inside the live browser and press Done, answer a one-time code or pick a sign-in method right in the card, or press Skip to let the model try another way. When the site is one of your website connections, the sign-in is kept for next time.

The browser has no security key and no passkey, and cannot use yours: it runs in the cloud with nothing plugged in and no phone nearby. A site that asks for one is steered to its other methods, an authenticator code (which Pad can type from a saved authenticator key), a prompt to your phone, a text, or an email, and the model asks you when the site offers none. If an account signs in only with a security key, add an authenticator app to it before connecting.

Signed-in sites

Connect a website once and the browser starts every chat already signed in, for every website connection you hold among the chat's sources. See Website connections.

Sessions, recordings, and cost

A chat has one browser at a time. It closes after ten minutes idle and reopens when needed; End session in the pane's menu closes it now. Sessions are recorded, a replay plus the step screenshots, unless a connection in use is marked sensitive. Each browser step costs credits at the rate the rate card shows.

Code the model runs in the chat's sandbox can drive the same browser, so a many-step task can run as a script while you watch.