Use computer_action when an agent needs to interact with the browser like a user: click, type, scroll, drag, copy, paste, or inspect the screen. Pass one action for a simple operation, or batch several actions into one call to reduce latency.
Include a screenshot as the last action when you need to see the result. screenshot, read_clipboard, and get_mouse_position return data, so they must be the last action if included.
Parameters
Action types
Recommended flow
- Start with
screenshot so the agent can see the viewport and coordinate space.
- Batch related pointer and keyboard actions together.
- Add
sleep between actions that trigger navigation, animation, or async UI updates.
- End with
screenshot, read_clipboard, or get_mouse_position only when you need returned data.
Search from the page
Clipboard example