Computer use

Two kinds of eyes

Computer use reads through two channels at once:

• The accessibility tree gives windows, controls, text, and hierarchy. It is precise, targetable, and tells us what an element is rather than what it looks like.
• Screenshots cover what the tree cannot express — canvases, charts, and custom-drawn interfaces.

Together, the agent knows both the control’s name and its position on screen.

Background input

Actions are delivered as background mouse and keyboard events:

• Your pointer does not move, so you can keep working.
• The target window is not raised — input lands even when the window is fully covered.
• For the rare app that truly needs foreground input, there is an explicit opt-in fallback.

Controls, not coordinates

Because targeting is based on the control itself:

• Moving a window, changing resolution, or changing display scaling does not break an action.
• A localized build of the same interface is still recognized.
• Coordinate input remains available for canvases, games, and handwriting.

Safety boundaries

• The agent states what it is about to do; irreversible actions need your confirmation.
• Accessibility and screen recording permissions are granted by you and can be revoked in System Settings at any time.
• It does not download or install software on its own; required apps come from you, from official channels.