Show HN: Dayflow – A git log for your day (github.com)

Dayflow is a native macOS (SwiftUI) app that turns your screen activity into a visual timeline with AI-generated summaries, distraction highlights, timelapses and a daily journal. It records the screen at 1 FPS in 15‑second chunks, analyzes recent footage every 15 minutes, and auto-deletes recordings after 3 days. The app is compact (~25 MB), uses ~100 MB RAM and under 1% CPU, integrates Sparkle for updates, and is fully open source (MIT) so users can inspect exactly what “watching your screen” does. Its core significance to the AI/ML community is the privacy-by-design, provider-agnostic approach and the practical tradeoffs between cloud and local inference. You can point Dayflow at Google’s Gemini (BYO API key) for efficient native video understanding (a two-LLM-call pipeline), or run local servers (Ollama/LM Studio) which reconstruct video understanding via many frame-level calls (33+ LLM calls), yielding higher latency, more compute and battery/GPU load but keeping data on‑device. The repo documents data paths, purge/reset steps, and nuances about Gemini’s “Paid Services” data‑use rules. For researchers and builders this is a useful, transparent reference for lightweight continual screen capture, video-to-text pipelines, and the practical costs of moving inference on‑device vs. cloud.