Fully local, fully private
whisper.cpp for speech recognition, Kokoro-82M on the Apple Neural Engine for the voices, Llama 3.2 3B via Ollama for summarising. No API keys, no per-token cost, no rate limits.
NovaHUD is a floating, always-on-top voice HUD for driving running Claude Code sessions hands-free — and the entire voice pipeline runs on your own Mac.
Free for 3 days, no card required · Requires macOS 14+ and Apple silicon
NovaHUD is a paid macOS app that gives you hands-free voice control over your running Claude Code sessions. It puts a floating, always-on-top card over every other window — including the Claude Code app itself — one card per session, permanently bound to that session.
You speak to a session by name, nickname, session title or orb colour ("blue orb, are you there?"), or say "everyone" to reach all of them at once. Each session answers in its own distinct human voice.
Everything runs locally. Speech recognition is whisper.cpp, the voices are Kokoro-82M running on the Apple Neural Engine, and the summarising and routing model is a local Llama 3.2 3B via Ollama. There are no API keys, no per-token costs and no rate limits, and nothing you say is sent to a server.
It needs macOS 14 or later on an Apple silicon Mac. It costs $5 a month ($3 a month with a referral code), after a 3-day free trial with no card required.
Every session gets its own card, permanently bound — drag it wherever you like and it will never quietly jump to another session.
NovaHUD routes what you said to the right session and the card updates live. No clicking into a terminal, no hunting for the right window.
Talk over a reply and it stops instantly, mid-sentence — the same way interrupting a person works.
A small, focused set of controls built for one job: keeping your hands off the keyboard while Claude Code works.
whisper.cpp for speech recognition, Kokoro-82M on the Apple Neural Engine for the voices, Llama 3.2 3B via Ollama for summarising. No API keys, no per-token cost, no rate limits.
A floating card sits above every other window, including the Claude Code app. Each one is permanently bound to a single session and never gets reassigned.
Every session gets a distinct orb colour and a distinct human voice. Address one by name, nickname, session title or colour — "blue orb, are you there?" — or say "everyone".
Read aloud speaks everything back. Summary speaks the bottom line and attributes it by session name. Quiet stays silent but still delivers what you say.
Start talking while it's mid-sentence and it stops, immediately. You never have to wait for a reply to finish before you redirect it.
Tiny, small, medium or large, and you pick the default. A tiny card is the orb and a status dot; a large one carries the full status line and every control.
NovaHUD runs the whole voice pipeline on-device. There is no server in the loop for anything you say, so there is nothing to leak, nothing to meter and no key to rotate.
api.novahud.app, for licensing, the trial clock and sign-inIt runs on your own Mac and there are no API bills behind it, so there is no reason to rent it to you forever.
The whole app, for three days
Everything, for the price of a coffee
Same app, with someone's code
What the thing actually looks like on a desktop, and an honest account of where it is the wrong tool.
The floating card, its four sizes, the eight session orb colours and what each listening mode does — drawn at full size rather than described.
Against typing at the terminal, against dictation, and against a cloud voice assistant — including the four cases where you should not buy it.
Three days, every feature, no card. Requires macOS 14 or later on an Apple silicon Mac.