01 / Give the job its own profile
Use a dedicated named profile and sign into only the accounts this task needs. Persistent profiles retain browser data across runs, although sites can expire sessions.
gripui attach workA profile keeps browsing contexts separate. It is not a sandbox for an AI agent that also has access to your files or shell.
02 / Let a person sign in
Have the agent navigate to the site. When a password, CAPTCHA or sign-in verification needs a person, hand the browser over:
gripui human-drivesComplete sign-in yourself in the visible window. Keep passwords and authentication codes out of the agent conversation.
Handoffs can restart the browser connection and restore tabs. A site’s unsaved form state may not survive. Authenticate first, then start the detailed work.
03 / Hand it back and look again
Once sign-in is complete, return control and refresh the page view:
gripui agent-drivesgripui hudConfirm that the site shows the signed-in state before continuing. Site restrictions still apply; a local browser does not guarantee access.
04 / Keep consequential choices explicit
Tell the agent the limits of the task: what it may change, what it may only inspect, and when you want to review. A signed-in session carries the account’s authority even if the agent never saw the password.
If a signed-in site asks for a device approval such as Touch ID mid-form, approve the prompt directly and have the agent refresh its view. A full handoff may lose the form you just prepared.
For a purchase, review the details and make the final commitment yourself. After any important change, verify the resulting record or page before calling the task complete.
See a real deployment workflow ↗