中文

JarvisVoiceHands-free voice typing and app control

A macOS menu bar app. It listens on its own and types what you say into the app you're using; say a command and it operates the app for you. Works with Claude, ChatGPT, your browser, LINE, Notes and more. It only types, it never sends; press Enter yourself.

Install from GitHub Architecture and numbers

Three layers, all on by default

Each layer builds on the one before; you can turn layers 2 and 3 off from the menu.

Layer 1

Voice typing

Finish a sentence (a pause of about 0.85 s) and the transcript is pasted where your cursor is. No keys to press.

Layer 2

AI cleanup

Finish a passage and the AI, looking at the text around your cursor, what's on screen, the app and the URL, turns it into what you meant to type: filler removed, self-corrections resolved, lists formatted, tone matched to the app, your own terms spelled right. Choose OpenAI, DeepSeek or Claude.

Layer 3

Auto-focus and voice commands

It writes even when your cursor isn't in a text field, by picking the right one. Say "open the TSMC stock page" and it clicks through the screen step by step. Each step is decided by Jev in about 0.2 s.

What it looks like

You say

Um, so we're meeting tomorrow at three, no wait, four, and, uh, bring the laptop

It types (after AI cleanup)

We're meeting tomorrow at four. Bring the laptop.

You say (cursor not in a text field)

Open the TSMC stock page

It does

Types "TSMC" into the search box
Presses Enter
Clicks "TSMC 2330" in the results
Sees the stock page and stops

The command example was measured on a test stock-app window: four steps in 3.3 s, 0.17–0.28 s per decision.

Set up in two steps

Step 1: install and open the app

You need macOS, Node.js 22 or later, and ffmpeg (brew install node ffmpeg).

git clone https://github.com/urusaqqrun/jarvis-voice.git ~/jarvis-voice
cd ~/jarvis-voice
./install.sh openai

To pay nothing for transcription, make the last line ./install.sh local (on-device recognition; your voice never leaves your Mac). The first time, allow two permissions: Microphone and Accessibility.

Step 2: ask your AI to set it up

Open the AI you use (Claude Code, Codex, ChatGPT, …) and paste this:

Please follow ~/jarvis-voice/AI_SETUP.md to set up JarvisVoice for me: fill in my vocabulary and run the tests, and ask me whether I want AI cleanup, auto-focus and voice commands; if I do, set them up and check that they work.

It finds the terms you use most in what you've typed before, asks which AI provider you want, has you paste your key into a file only you can read, and runs a check to show it works.

What it costs

Item	Rate	Talking 1 hour a day, per month
Transcription (OpenAI `gpt-transcribe`)	US$0.0045 per minute of audio sent	US$7.4
Transcription (on-device)	Runs on your Mac	Free
AI cleanup (DeepSeek, Claude or OpenAI)	About US$0.00003–0.00008 per passage	Under US$1
Layer 3 decisions (Jev, optional)	About US$0.00001 each	About US$0.1

You use your own keys and pay each provider directly. It's always listening, so people talking nearby get transcribed too; choose "Pause listening" in the menu during meetings or videos.

Where your data goes

Audio: only segments with speech go to the transcription service you chose; with on-device recognition it never leaves your Mac.
With AI cleanup on: 400 characters on each side of the cursor, the selected text, the last ~1,500 characters visible in the window, the app name, the window title and the URL go to the AI provider you chose.
Layer 3: the names of the buttons, links and text fields on screen go to the decision maker (Jev or your AI provider).
History, vocabulary, settings and keys stay on your Mac.