Browser mode

Browser mode gives a task a real browser: dashboards without APIs, supplier portals, admin consoles, anything with a login page. It’s built around one principle — the agent can use your credentials, but can never read them.

The vault

Credentials live in the browser vault, encrypted, and are injected into login forms without ever entering the model’s context. Every use is written to an audit ledger: which credential, which site, which session, when. You can review the ledger any time and revoke a credential in one click.

Grant a credential once; every future session that needs that site asks the vault, not you.

The live view

A running browser session shows a live view of the actual page. Watch every click and keystroke, or take over — the moment you interact, the agent yields until you hand control back.

Paused for authentication

Note

Some moments demand a human: captchas, two-factor prompts, first-time logins you haven’t vaulted. The session enters paused for authentication and asks you to log in through the live view. What you type there goes to the website only — the agent never sees it. When you’re done, work resumes.

Approvals still apply

Submitting orders, changing account settings, sending anything — Browser mode stops for approval exactly like Computer mode does. The approval card shows what’s about to happen on which site.

Failure and recovery

• Site changed and a step can’t find its target → the session shows the last page state; guide it in the session or take over and do the step by hand.
• Login loop → check the vault entry is current; re-authenticate once via the live view and update the credential.
• Session seems stuck → open the live view first. The answer is almost always visible on the page.

Good first tasks

• “Log in to [portal] and download this month’s invoices to my Library.”
• “Check stock and price for these three SKUs on [supplier site] and update my procurement table.”