Desktop Agent Uses Coordinate Grids for Visual Grounding

Feisty-Gas9764 · reddit · 2026-07-19

A desktop Agent prototype named Rosply avoids DOM scraping, opting instead for a visual grounding approach using "screenshots + coordinate grids" for desktop automation.

Its workflow is: Screenshot → Visual reasoning → Mouse/keyboard actions → Repeat. The author also incorporated loop detection and coarse-to-fine grounding to improve clicking accuracy. The project is source-available and supports local execution.

Original post →

More from coding & agent

coding & agent channel →