I run a vision model on every screenshot, locally, on a 4GB GPU

This project enables running a vision-language model locally on a 4GB GPU by processing every screenshot taken on a desktop, allowing for real-time, privacy-preserving AI analysis of user activity. It is designed for developers and power users who want to automate workflows or gain insights from their screen without relying on cloud APIs. Its interest lies in proving that sophisticated multimodal AI can run on consumer-grade hardware, democratizing access to local vision-based agents.