Shipped: MicroLog
Most of what we read online disappears by lunch. Articles, release notes, random threads... you read them, but saving them is too much work, so you just move on.
I built MicroLog to fix that.
It sits in your Mac menu bar and silently tracks what you actually read. When you dwell on a page for a bit, it automatically creates a clean note with a summary, key points, and tags. Zero typing required.
Under the hood:
Qwen3.5 (0.8B) generates the notes right on your machine using llama.cpp.
EmbeddingGemma 2 acts as the memory, handling deduplication so you don't end up with repeat notes for the same topic.
It has a simple dwell filter so quick app switches don't trigger anything, and private apps like Mail, Terminal, and password managers are completely ignored.
Everything runs 100% on-device. No accounts, no cloud APIs, no network calls after the initial model download.
It's free, open-source (MIT), and built for Apple Silicon Macs. Since this initial release isn't notarized yet, you'll just need to do the usual right-click "Open" trick the first time.
Download: github.com/abhishek085/microlog
It's a v1, so it definitely has rough edges. Try it out on your daily reading and let me know where it breaks or what can be better.
#localAI #openSource #productivity #reading #macOS






Comments2
Running Qwen 0.8B through llama.cpp for this is a smart call, it keeps everything private. How's the battery and fan impact when it's summarizing all day? I build Android and Wear OS apps, and on-device models there are a lot harder with background limits, so I'm curious what dwell time you landed on.
Welcome to Lounge, Abhishek. Dedup with embeddings is the part I'm curious about: how did you set the similarity threshold so a follow-up article on the same topic becomes a new note instead of getting merged? And does a 0.8B model hold up on long, messy pages?
2 comments
Running Qwen 0.8B through llama.cpp for this is a smart call, it keeps everything private. How's the battery and fan impact when it's summarizing all day? I build Android and Wear OS apps, and on-device models there are a lot harder with background limits, so I'm curious what dwell time you landed on.
Welcome to Lounge, Abhishek. Dedup with embeddings is the part I'm curious about: how did you set the similarity threshold so a follow-up article on the same topic becomes a new note instead of getting merged? And does a 0.8B model hold up on long, messy pages?