Skip to content

How do I sign in to a website during a browser task?

You sign in yourself, never the assistant. It opens the site's sign-in page and hands you the keyboard. Your password is never stored, and later runs on that site start already signed in.

Last updated September 01, 2026

On this page

You sign in yourself. The assistant never signs in on your behalf, and it will not type a password, a passcode or a security code for you — not on any site, and not if you ask it to.

What it does is get you to the door. It opens the browser directly on the site's sign-in page, hands you the keyboard, and waits.

How it goes

  1. Ask for something on a site you need to be signed in for, or ask it to sign you in to a site.
  2. The assistant proposes the sign-in and a confirmation card appears in the conversation, with Cancel and Approve. Nothing opens until you press Approve. This is the one browser thing that always asks first — an ordinary browser task just gets going.
  3. The Computer panel opens on that site's sign-in page. In the conversation, the card for it reads "Sign in to continue", and its button reads Open and sign in if you need to get back to it.
  4. An amber strip across the top of the panel says: "Sign in below and the session is saved for future runs. This window isn’t recorded, and your password is never stored — only the signed-in session is kept."
  5. Type your details in yourself, including any two-step code. The assistant is not driving during this. A sign-in window is never assistant-driven; you have the wheel for all of it.
  6. When you are through, press I’m signed in.
  7. You will see "Saved. Future runs will start signed in." If something went wrong you will see "I couldn't save that sign-in — nothing was stored. Try signing in once more." Do exactly that: nothing was stored, so there is nothing to undo first.

Why it works this way

Two reasons, and the second is the one that matters to you.

The first is that the assistant is not permitted to type a password, a passcode, or a security, one-time or two-step code. That is not a rule it is asked to follow and might get talked out of — it is enforced underneath it. The action is refused and the run stops there. Asking again in different words does not change the answer.

The second is that your password never has to go anywhere near this system. You type it into the site, exactly as you always would. That sign-in window is not recorded, and the password is not stored. What is kept is the signed-in session itself, so the next run on that site starts already signed in.

Note: you never need to give the assistant your password, and you should not paste one into the chat. There is no request that makes it type one.

After that, you stay signed in

Once a sign-in is saved, later runs on that site start already signed in. That holds when the browser sleeps and wakes, and across conversations — it comes back knowing who you are.

Saved sign-ins are per site, and you name them, so you can tell them apart. Ask the assistant to list them and it will tell you which are ready to use and which have gone stale.

When a sign-in goes stale

They expire on their own, and there is nothing wrong when they do. Changing your password does it. So does "log out of all devices". So does a site simply timing the session out.

When a run finds itself logged out, it says so and asks you for one more sign-in from the panel, rather than retrying blindly. Sign in again exactly as above and it carries on from there.

One more thing: a saved sign-in cannot be used by two runs at once. If a second run wants it, it is told to try again in a minute — that is deliberate, because using it twice over risks the site logging you out.

If a site asks for a code partway through a task

The same rule applies mid-task. If a site throws up a two-step code, a password prompt or a "confirm you're human" check in the middle of an ordinary browser task, the assistant stops right there. The browser stays open on that page, and it tells you what it stopped at and asks you to do that step.

Take the controls (clicking anywhere on the live view does it), do the step, then hand back. The assistant reads the page again and carries on from where you left it rather than starting over. None of this is recorded as a failure.

What it will never type, sign-in or not

  • A password, a passcode, or a security, one-time or two-step code.
  • A payment card number.

And if it cannot tell what a field is for, it refuses to type into it at all rather than guess. In every one of those cases it stops, leaves the browser open on the page it reached, and asks you to fill that part in yourself.

Was this page helpful?
Still need help? Contact support