Read. MicroLog takes the notes.
It sits quietly in your menu bar and notices what you've actually spent time reading — an article, a doc, a thread. No highlighting, no clipping, no typing. Everything runs on your Mac, so nothing you read is ever sent anywhere.
AI models download once, then it works offline
You read · 4 min
A long piece on vector databases and how embedding similarity is used for semantic search — you stayed on the tab well past the dwell threshold, so MicroLog quietly picked it up.
MicroLog wrote
Vector databases, explained
- Embeddings place similar meaning close together in vector space
- Approximate nearest-neighbor search trades a little accuracy for speed
- Indexing strategy matters more than raw dimension count
How it works
No clipping, no account, no cloud. Just read — the note appears on its own.
Capture
Tracks the frontmost app/window via Accessibility, with a ScreenCaptureKit + Vision OCR fallback when there's nothing else to read.
Filter
Only counts it once you've dwelled ~20s and ~150 words — quick app switches never create noise.
Dedup & thread
URL canonicalization, content hashing, SimHash and embedding similarity catch duplicates and link related reads into one thread.
Summarize
A local LLM fills in the title, summary, key points, "why it matters" and tags in the background.
The AI models it uses
Downloaded once on first launch. Nothing is sent to a server after that.
| Job | Model |
|---|---|
| Summarization title, summary, key points, tags | Qwen3.5 0.8B Q4_K_M, via llama.cpp |
| Dedup & similarity catches rereads and related threads; also embeds images | EmbeddingGemma 2 UD-Q4_K_XL + mmproj, via llama.cpp |
Both run locally over llama-server on localhost — no network calls beyond the one-time model download.
Download for Mac
Or build it yourself from source:
Download & open
Get MicroLog.dmg, open it and drag MicroLog.app onto the Applications shortcut.
Open it the first time
Not notarized yet, so Gatekeeper will block it — right-click the app → Open → Open again. Just once.
Grant permissions
Accessibility, Screen Recording, Automation and Notifications when asked.
Let it set up
MicroLog downloads its AI models (Qwen3.5 0.8B, EmbeddingGemma 2) on first launch, then runs fully offline.
Private by design
- No account, no cloudThere's nothing to sign up for and no server that sees what you read.
- Everything on-deviceCapture, OCR, and summarization all run locally — no analytics, no cloud APIs.
- Denylisted apps are skippedMail, Messages, Terminal, and password managers are never read.
- Open sourceMIT-licensed. Read every line on GitHub.
Questions
Do I need internet?
Only once, to download the two AI models on first launch. After that MicroLog works fully offline.
Does it read everything I look at?
No. It only turns a window into a note after you've dwelled on it for ~20 seconds and it has ~150+ words, and it skips denylisted apps like Mail, Messages, Terminal and password managers entirely.
What happens if I reread the same article?
The dedupe engine recognizes it via URL canonicalization, content hashing, SimHash and embedding similarity, and skips creating a duplicate note.
Does it work on Intel Macs?
MicroLog targets Apple Silicon Macs.
Is it really free?
Yes, and open source under the MIT license.