Skip to content

Features

A-Eye

A-Eye lets you click an element on the page to make it the subject of your next message, and lets the agent ask you to point when it cannot tell which element you mean.

3 min read Edit this page on GitHub

Pointing at something

Click the A-Eye button (the lens in the side panel's composer row). Your cursor becomes a lens and outlines whatever you hover. Click to pick it. ↑ widens the pick to the parent element when you land inside something smaller than you meant; Esc cancels.

The pick shows as a chip above the composer, with a small thumbnail, and goes out with your next message: the element's tag, role and selector, the text it held at that moment, and a photograph of it as you saw it. The chip clears when the message sends, so one pick scopes one message. The chip's × drops it without sending.

The site never sees the pick. Every pointer event in the sequence is stopped before the page gets it and the click itself is cancelled, so picking a "Delete" button does not press it.

What the agent does with it

The element arrives in the run's system prompt as the subject of the instruction. Ask "what does this say?" and the answer is about that element; say "translate this" and only that is translated. The agent re-reads the element live before acting on it, because the page can change between your pick and its first tool call. When appearance matters, it can view the photograph taken when you picked, even if the page has changed since.

Your words win when they clearly point elsewhere: "now go to checkout" has left the element behind, and "compare this with the one below" is about two elements. The pick resolves ambiguity; it does not override what you say.

If the element is gone by the time the agent looks, it says so instead of acting on whatever is nearest.

When the agent asks you to point

page_pickElement opens the same lens from the agent's side. When a target is ambiguous (three rows share a label, or "the second one" could mean two things), the agent can hand the page back to you with a one-line question over it, and continue with whatever you click.

It stops the run and waits for a person, so it is the most expensive tool in the set, and the a-eye skill tells the agent to use it rarely. Dismiss it with Esc and the agent is told to ask in words instead of asking you to point again.

Edges

  • One tab. The lens opens on the tab that was in front when you pressed the button. Switching tabs mid-pick leaves it on the old tab; it times out after a minute.
  • One lens at a time. Starting a new pick, from the button or from the agent, dismisses a lens already waiting, and whoever opened that one is told it was cancelled.
  • Top frame only. An element inside an iframe picks the iframe, not what is in it.
  • Never Browsentic's own. The hands-free mic, the minimized rail and Browsentic's cards are skipped: the lens draws nothing over them, and a click on one picks nothing. A pick started from the hands-free mic's menu moves the mic aside until it is done.
  • The element, not the page. The photograph covers what you picked plus a small margin. For the page around it, ask for a screenshot as well.
  • Long elements are cut. Picking a whole article sends the first couple of thousand characters; the agent reads the rest with page_extractText scoped to the selector.

See also