Other notetakers upload your audio. Minutiae won't.
Records both sides of the call, transcribes and writes the note on your Mac. If you've used Granola, it's that, minus the upload.
No bot in the room
Minutiae records your mic and your Mac's audio. Nothing joins the call. Works with any call app.
No server hears you
Signed out, once the models are on disk, it makes zero network requests. Watch for yourself:
nettop -p minutiaeLittle Snitch or LuLu show the same. Sign in and it sends text, never audio.
No lock-in, ever
Notes are Markdown files in your folder. The code is MIT licensed. Nothing can be taken away later.
The difference
Where your meeting goes.
Most notetakers promise to protect your recording. Minutiae has nowhere to send it.
A typical AI notetaker
- 1 Your microphone
- 2 A bot joins the call and records everyone
- 3 Audio is uploaded to the vendor's cloud
- 4 A third-party model provider transcribes it
- 5 A summary is generated by another model
- 6 Everything is retained under their policy
Six places your meeting exists. You control one.
Minutiae
- 1 Your Mac
Everything below happens inside it
- Listen Your mic, and whatever comes out of your speakers
- Transcribe Parakeet TDT v3 on the Neural Engine, via FluidAudio
- Write Qwen3.5-4B, 4-bit, on the GPU through MLX
- Save Markdown and audio in a folder you picked
One place your meeting exists. Nothing else to audit.
Turn on sync and the text gets a second home on your other Mac. The recording stays put. How sync works.
How it works
Record. Read along. Enhance.
One window. A live transcript, and your notes beside it.
Press Record.
Your mic and your speakers, recorded straight from the Mac. Nobody sees a new participant.
Two channels, not diarization.
Your microphone is one column, everything your Mac plays is the other. It is not per-person: in a hybrid meeting, a shared room mic arrives as one voice.
Jot while you listen.
The scratchpad saves with the session. Your thoughts survive the meeting.
Enhance, and it's a note.
A local model turns the transcript and your notes into a finished Markdown file. Obsidian opens it as-is.
No hallucinated text
Silence produces nothing.
Quiet audio never reaches the model. A pause stays a pause, not a sentence nobody said.
Nothing leaves this tab. Check your browser's network panel.
Transcript
Nothing yet.
What a transcriber invents during a pause.
Local-first
Your audio never leaves your Mac.
No telemetry. No account unless you want one.
Open the app and press Record. Nothing reports back. Sign up only if you want your Macs in sync.
The models, by name.
Speech is Parakeet TDT v3 (0.6B, 461 MB) on the Neural Engine through FluidAudio. Notes are Qwen3.5-4B at 4-bit on the GPU through MLX. Neither needs a connection once downloaded.
Recordings and notes are plain files.
Audio, transcript and note live in folders you can open, move, back up or delete. No database owns them.
Every network request the app makes
Idle memory. The note model loads only while it writes.
From pressing Record to capturing audio.
Transcript written during silence.
Of transcription, on your Mac. Always.
Measured on an M1.
Optional
Three ways to run it.
The transcript is the meeting, so cloud is off until you turn it on. Text only. Audio never moves.
Local. Free.
No account. Nothing leaves your Mac. Notes are unlimited and always will be.
Cloud. Free.
Sign in and your transcripts and notes sync across your Macs. A hosted model writes three notes a month.
Cloud. $5 a month.
Same sync. Unlimited hosted notes, within fair use. Switch back to the local model any time.
What the cloud tiers send
Fair use is the paid tier's only limit. Past about fifty hosted notes in a month we may use a lighter model or slow down, and the app says so when it happens.
Side by side
Against the other local-first apps.
These are the real alternatives. They run on three platforms and take any OpenAI-compatible endpoint. Minutiae doesn't, yet. Here's what it does instead.
| Capability | anarlogwas Hyprnote | Meetily | Minutiae |
|---|---|---|---|
| Platforms | macOS, Windows, Linux | macOS, Windows, Linux | macOS only, Apple Silicon |
| Built with | Tauri 2, React | Tauri, Next.js | Tauri 2, Svelte 5, Swift audio engine |
| Notes stored as | SQLite, Markdown on export | Local database | Markdown files, authoritative |
| Transcription | On-device on some Macs, else a hosted provider | Whisper or Parakeet, local | Parakeet on the Neural Engine, always local |
| Note model | Claude via OpenRouter by default | Ollama, or a hosted provider | Qwen3.5-4B on your GPU; hosted is opt-in |
| Bring your own endpoint | Yes — any OpenAI-compatible server | Yes — any OpenAI-compatible server | Not yet |
| Speaker separation | Diarization | Diarization | Two channels only — you, and everyone else |
| Price | MIT; enterprise tier is commercial | Free; PRO tier is paid | MIT; cloud free, or $5/mo unlimited |
Straight answers
The questions you were going to ask.
Is it signed and notarized?
Yes. Developer ID, hardened runtime, notarized and stapled. Not App Sandboxed — a process audio tap can't be.
Electron?
No. Tauri 2, so the UI is the system WebView. Audio and speech recognition are a separate Swift binary. 20 MB download, 64 MB installed.
What does it cost you at rest?
Under 250 MB idle with the speech model loaded. Capture starts in under a second. The note model loads only while it writes, and that is the peak.
Windows or Linux?
No. The Windows sidecar has an IPC seam and no capture. Per-process audio tapping is a different problem on each platform. macOS 14.4+ on Apple Silicon.
Can I point it at my own model?
Not yet. An OpenAI-compatible endpoint for Ollama and LM Studio is the most-asked-for thing, and it isn't built.
Is it legal to record?
Depends where everyone is. Some places need every party's consent. The app tells no one it is recording. Know your local rule.
Your next meeting can stay yours.
Free, open source. No account needed.