Cobinar Lens checking edge

Read any region of the screen.

Drag the reticle over a chart, a label, a paragraph, anything. The crop is stripped down to palette, geometry and text lines in the browser, and only that summary travels — the model never sees your pixels.

Sample dashboard with a bar chart and two figures
Try it here. Box the bar chart, then box one of the dark tiles.

What comes off the crop

Palette
up to six merged colours with their share of the frame
Geometry
connected components typed as rectangle, ellipse, line or blob
Text lines
glyph clusters grouped onto baselines, with boxes
Luminance
a 12×12 grid standing in for the picture itself
Stats
brightness, contrast, colourfulness, sharpness, edge orientation
Hash
64-bit difference hash for matching repeat captures

Where a capture goes

  1. The browser crops the region you boxed, from the page image when it can and from a tab capture otherwise.
  2. lens-extract reduces it to a few kilobytes of structured JSON. Nothing leaves until this finishes.
  3. The edge worker runs the plugin pipeline over that JSON and signs the call to Cloud Run.
  4. Groq reads the descriptor and answers in text.
  5. The panel shows the reading and pulls related headlines straight from public feeds.