Hold a key and talk. Amanu writes your words at the cursor, in any app — transcribed entirely on your Mac.
fn is the whole interface.
Push-to-talk, like a radio. Hold the key while you talk; your words land at the cursor the moment you let go. Bind any key, hold or toggle.
fn by default · rebind any key · hold or toggle
Two ways in: native Accessibility insertion where apps support it, paste-swap where they don't. Your clipboard — images included — is restored either way.
AX insertion · paste-swap · clipboard preserved
Parakeet transcribes on the Neural Engine; Refine runs on the GPU. In memory, nothing uploaded.
Parakeet TDT v3 · CoreML · FluidAudio
An optional on-device pass over your words. Raw keeps every one. Punctuate adds punctuation and changes nothing else. Clean drops the filler and tidies grammar — a guardrail keeps it from inventing words.
per-app rules: terminal → Raw · Slack → casual · Mail → formal
While you speak, a small capsule shows the waveform — it never steals focus. When you stop, it leaves.
capsule · 220×44 pt · VoiceOver and Reduce Motion aware
Menu bar only — no Dock icon. Quick actions for dictation, Refine mode, and recent transcripts. Launch at login; one process, no daemons.
menu bar app · SF Pro · light and dark
Say “new line”, “new paragraph”, “scratch that”, “all caps”. Commands are parsed deterministically — never guessed by a model.
parsed, not guessed
Transcripts are encrypted at rest, the key in your Keychain. Keep them a day, forever, or not at all — zero-retention stores nothing. Search, copy, delete.
AES-GCM · Keychain · zero-retention mode
Speech recognition is NVIDIA Parakeet, run on the Neural Engine through CoreML. The optional Refine pass is an open language model — 4-bit, run on the GPU through MLX. No API key, no endpoint: the app pulls the weights down once, and dictation never touches the network.
This is the difference from built-in dictation: its model is sealed inside the OS. Amanu's are files you choose — pull a faster one, or a more accurate one, the day it ships.
Parakeet (~600 MB) downloads on first run; Refine models on demand. Checksum-verified, over HTTPS, once.
Models sit in Application Support like any other file. Inspect them, delete them, pull them again.
Parakeet on the Neural Engine, Refine on the GPU. The network is not part of the path.
There is no server to send your voice to. Transcription happens on the Neural Engine, in memory, while you hold the key. We built it that way so the promise is structural, not policy.
No telemetry by default. The network is used for three things: model downloads, update checks, and crash reports you opt into. History is encrypted with a key that never leaves your Keychain — or kept not at all.
A low-level tap listens for your push-to-talk key — only that key, only while held.
Places your words at the cursor in the focused app.
Detects secure input and will not type into password fields.
No screen recording, ever.
In Rome, the a manu was the secretary who stood at your side and wrote down what you spoke — the amanuensis, first set down in full by Suetonius. So Amanu isn't a clipping that happens to sound nice; it's the actual root. For an app that lives under a key your finger rests on, at hand is almost suspiciously perfect — and the app named by hand exists to remove the hand entirely. Your hands stay on the keyboard, the wheel, the decks. The writing happens anyway.
The mark in your menu bar is a Tironian et — from the notae Tironianae, the shorthand Tiro invented when speech outran his stylus. It outlived Rome by a thousand years; a fragment still marks Irish road signs today. Cicero's amanuensis was the original speech-to-text engineer.
The thinking stays yours. The transcription disappears into something trusted. The work leaves your head faster than hands ever could.