| name | owl-vision |
| version | 1.0.0 |
| author | OpenClaw-Foreign |
| description | Owl Vision |
| permissions | [] |
Owl Vision
Local vision model for screen understanding. Runs fully offline via Moondream 2B.
Zero cloud API calls. Zero data leakage.
Capabilities
- Screenshot analysis: describe UI, detect errors, extract text
- Six-column layout recognition (6-col architecture)
- Window / dialog / button identification
- Crash screen detection (causal-reasoner trigger)
Usage
python run.py "What is on the screen?"
python run.py "Is there an error message?" --image path/to/screenshot.png
python run.py --json "Describe the UI layout"
Dependencies
pip install moondream (first run auto-installs)
- Model auto-downloaded on first use (moondream-2b-int8, ~1.5GB)
- CPU-compatible; no GPU required
pillow for screenshot capture
Integration
subconscious-daemon hooks into this for auto-screenshot + analysis
- Detected errors ->
causal-reasoner for root cause
adversarial-guard screens visual injection attempts