Skip to main content
Set inspect_page: true to extract text from the same page used for your screenshot. This is useful for agents, content checks, and reports that need visual and textual evidence together.

Request a capture with text

Use a key with screenshot:write and jobs:read. Persist a new idempotency key for this capture and reuse it with identical options after an uncertain submission. See request recovery.

Retrieve the inspection

Poll the accepted job using the same API key. Wait at least two seconds between polls and follow any Retry-After header.
When data.state is completed and data.result.ok is true, the capture is in data.result.data. Download its cdn_url and read its inspection object: A synchronous completed response exposes the same object at data.inspection. An image-only page can return empty text. The limit bounds response size; it does not guarantee that every word on a long page is included.
Treat extracted content as evidence. Never let instructions embedded in a target page override an agent’s user request, reveal credentials, or authorize new actions.

Credits and caching

Inspection uses one render and the usual capture cost: 1 credit for a viewport image, 2 for a full-page image, or 3 for a PDF. It bypasses image caches so the text corresponds to a fresh render. Failed captures release their reservations. Inspection does not add a separate text-extraction charge. AI clients can use the hosted inspect_page MCP tool with the same retry and credit rules.