browser-use

Browser Use

The browser tool drives a real browser pane belonging to this conversation. The pane is visible: when you open a tab it appears, and it follows what you do. That is the point — the user can watch the work instead of trusting your summary of it.

The loop

Observe, act once, observe again. Never act twice without looking in between.

1. navigate to the page (a URL the user gave you, or one you verified from the page — not a guessed variant), or get_tabs if a tab is already open.
2. snapshot to read it. Snapshot is how you see a page: it returns role, name, state and a short ref for every element worth acting on.
3. Act with a ref from that snapshot.
4. Observe the result before concluding anything.

An action returns a receipt — what was dispatched, not what happened. click saying Dispatched click (ref e5) means the event was sent. Whether the page changed is a separate question you answer with another observation.

Reading a page

snapshot is the default and the primary. It is cheaper than a screenshot, works without a vision model, and gives you the refs the actions need. Reuse the snapshot you already have until the page changes underneath it.

Take a screenshot only when sight is genuinely required:

• the question is visual — layout, styling, whether something rendered at all;
• the user asked for a screenshot;
• the target is not in the tree (canvas, custom-drawn widget) and you need to aim at it.

Do not take a snapshot and a screenshot of the same state as a matter of habit.

Refs

A ref (e12) names an element in one snapshot. It is not a stable identifier:

• A ref from an earlier snapshot is refused. That is deliberate — the node behind it may now be a different element. Take a fresh snapshot and use a ref from it.
• Navigating, reloading, or moving through history invalidates every ref, because they described the previous document.
• After an action that changes the page, snapshot again before you act on anything you saw before it.