Vision playground

About this playground

OmniParser V2 turns a screenshot into structured data for computer-use agents: bounding boxes for interactive elements, recognised text, and an annotated image. Adjust the detection thresholds in the panel and re-run to see how they change the result; the same parameters are accepted by the API.

Reading the output

Each detected element comes back with a type (icon or text), a bounding box in normalised 0–1 coordinates, whether it looks interactive, and a short caption or the recognised text. The annotated image draws numbered boxes so you can match ids to elements. Lowering the box threshold finds more, smaller elements at the cost of duplicates; raising the overlap threshold merges boxes that cover the same control.

Typical uses

Grounding a browser or desktop agent (find the "Submit" button, then click its box centre), auditing UI screenshots for accessibility text, and extracting labelled controls from screen recordings frame by frame. A request costs $0.003 regardless of image size, and images are not stored after the response is returned. Uploads here go through the same endpoint your code would call, so what you see is what the API returns.