Dayflow automatically records specific task details and time by analyzing screen content with VLM

While conventional tools only record which application was open, Dayflow understands the content displayed on the screen. According to the developer, the system uses a local or cloud-based VLM to analyze screenshots and create a semantic timeline based on the context.

This allows users to grasp specific work details and durations—such as "Reviewing PR comments: 45 minutes" or "Browsing Rust-related HN threads: 20 minutes"—rather than coarse records like "Chrome: 3 hours."

To protect privacy, the app runs a model called qwen2.5-vl-3b (approximately 4GB) locally by default. No cloud uploads are required, and no account creation is necessary.

The technical stack consists of SwiftUI and SQLite. The LLM engine is not bundled; instead, it is designed to integrate with existing tools such as Ollama or LLMStudio. The developer stated that they spent significant time on evaluation and prompt chunking to balance quality when migrating from large-scale models to local small-scale models.

Planned future features include improving accuracy through distillation using Gemini outputs as teacher data and implementing a custom dashboard that can answer arbitrary questions. The source code is released under the MIT license GitHub.