Skip to main content
Drive a browser session with raw mouse and keyboard input at screen coordinates. Pass a single action for simple operations, or pass multiple actions to batch them into one request for lower latency.
Prefer execute_playwright_code for anything a selector can reach. It’s faster, deterministic, and doesn’t depend on the model’s ability to locate a target in an image. Reach for computer_action only when there’s no selector to target: canvas apps, embedded PDFs, native dialogs, and drag interactions.
Coordinates come from a screenshot call, and are only as accurate as the calling model’s pixel grounding — models vary widely here. Take a screenshot after acting to confirm the result.

Parameters

Action types

read_clipboard and get_mouse_position return data, so they must come last. To capture the page, call the separate screenshot tool.

Example

Then call screenshot to see the result.