Capture a PNG screenshot of a browser session. This tool is read-only: it observes the session without changing it. Use it to see page state, confirm what an automation did, or diagnose a flow that’s stuck.
Your MCP client must pass image content from tool results through to the model. Most clients do, but a text-only model can’t read the screenshot even though the call succeeds.
Parameters
Response
Two content blocks: a text block describing the coordinate space, and the PNG image.
For a full-screen capture, image coordinates are screen coordinates, so you can pass them straight to computer_action:
For a cropped capture, the image starts at the crop, so add the offset before using a coordinate:
Example
Crop to a region:
To act on the page, prefer
execute_playwright_code — selectors are faster and more reliable than clicking at coordinates read off a screenshot.
