Open Source Transcription.Local or cloud.

A 1.9 MB menu-bar app for macOS. Hold Right ⌘ in any app, speak, let go — the text appears at your cursor. Run it on Gemini with your own key, or entirely on your Mac.

View on GitHub

Build it yourself in one command. Bring a Gemini API key — or install the local model and skip the key entirely.

New Message

Toteam@studio.co

SubjectThursday's release

Right ⌘

Small enough to read

A 1.9 MB binary built from about 5,400 lines of Swift. No Electron, no framework, no bundled dependencies — you can read the whole thing in an afternoon and compile it in one command.

Idle costs nothing

No timers, no background polling, no audio session sitting open. The microphone opens when you hold the key and closes when you let go. Menu bar only, no Dock icon, nothing to switch to.

Nothing to configure

Your language is detected automatically — German and English out of the box, on either engine, with nothing to switch. Punctuation comes back already in place.

Two engines. One switch.

Pick where the audio goes in Settings. Gemini keeps nothing on your disk; the local engine keeps nothing on the network. Either way the app itself stays the same small thing in your menu bar.

Cloud

Gemini 3.5 Transcribe

Streams while you speak, so the text is already waiting when you let go. Smart mode tidies a dictated list into real markdown.

On disk: Nothing — no model to download
In RAM: Nothing — the work happens elsewhere
Needs: A free Google AI Studio key
Audio: Goes to Google, and nowhere else

Local

Whisper, on this Mac

Whisper Small — the quantised ggml-small-q5_1 build — through whisper.cpp. No key, no account, no request. Transcribes after you release the key rather than while you speak.

On disk: 190 MB model, ~25 MB engine
In RAM: Nothing between dictations
Needs: One brew install, then one click
Audio: Never leaves the machine

The local model is downloaded on demand from the official whisper.cpp repository, checked against a pinned SHA-256, and removable from Settings in one click. One process handles one dictation and exits with it, so nothing sits in memory waiting for you to speak again.

Build yourself, stay in control.

The source is on GitHub and builds with one command. That way the app that types into every one of your windows is one you read and compiled yourself — and if you would rather it never phoned home at all, that is a switch, not a fork.

View on GitHub