Local screen ring for DeepSeek Harness: keeps the last ~45 seconds of screen frames in memory so the model can rewind what just happened, recalls any frame at 1:1 pixels, reads window text through the Windows accessibility API instead of OCR, locates controls by name and reports coordinates only (never clicks), and masks private regions in every capture. Windows only; everything stays on the machine. Its HTTP routes have no DSH access check, so a web page open in your browser could call them.
1vorgesternVision, Sprache & MultimodalMIT