Voice Control for macOS
AGI Controller
A voice controller for macOS that is always ready, acts before you finish the sentence, and does not guess.
What you say × what happens
Say it in English or Japanese. Examples:
"open Safari"
「Safariを開いて」
"click log in"
「送信を押して」
"type hello world into the search field"
「こんにちはと入力して」
"search for alan turing"
「東京の天気を調べて」
"scroll down a bit"
「一番下まで」
"put this on the left half"
「左半分にして」
"export as PDF"
「サイドバーを表示して」
"run my morning routine"
「ショートカットの日報を実行」
"no, not that one"
「違う、そっちじゃない」
Hold fn and speak
Rebind the trigger to any key, modifier key, or extra mouse button.
Hold fn
It listens only while you hold the key.
Double-tap fn
Starts a hands-free session. Double-tap again to end it.
Press Esc
Cancels listening.
It does not guess
When it is not sure, it checks before it acts.
Ambiguous target
Numbered badges appear on the candidates. Say "two" to pick one.
Destructive actions
Send, delete, buy, discard, and quit wait for a spoken "confirm".
Wrong action
"No, not that one" reverses the last action.
Chit-chat in a hands-free session
Talk that is not a command is ignored.
How it works
Hold fn and speak
It listens only while the trigger is held.
Transcribed on your Mac
Apple on-device speech recognition.
JEV answers with probabilities
About fifteen questions in a few hundred milliseconds.
Code decides
Act, wait, ask, confirm, or ignore.
A decision on every transcript change
Every time the partial transcript changes, one request goes to JEV, the System One model from TypeSafe. JEV answers about fifteen typed questions with calibrated probabilities in a few hundred milliseconds. Code decides what happens next: act, wait, ask which one, ask for confirmation, or ignore.
JEV never generates text
What to type, which URL, which app, and which on-screen control are candidates extracted by code. JEV only picks one, and the pick is copied verbatim. No generative AI is involved, so nothing you say is rephrased.
Only text leaves your Mac. Audio never does.
Speech is transcribed on your Mac with Apple on-device speech recognition, and no screenshots are taken. Requests go to AGI Backend with your sign-in and from there to JEV. The app holds no API key.
What leaves the Mac (text only)
- What you said, as text
- The name, window title, and URL of the frontmost app
- Up to 100 control labels and 150 menu command names
- Names of running, Dock, and mentioned apps
- Your Shortcut names
- The last three actions
- With AGI Cockpit installed: its task names and open choices
What never leaves the Mac
- Audio
- Screenshots
- The content of password fields
- Your clipboard
- Usage insights
Features
It reads the frontmost window through the macOS Accessibility API, so it works across apps.
Works across apps
Checked with Finder, Safari, Chrome, System Settings, and Electron apps such as Discord. It reads up to 100 control labels from the frontmost window.
Notch HUD
Live transcript, intent and target probability bars, numbered candidates, latency, and estimated cost. The probability bars can be hidden.
Dictation
Hold the trigger in a text field, and a sentence without a command is written word for word.
English and Japanese
Both the UI and speech. A second speech language can have its own trigger.
Shortcuts and menu commands
Runs your macOS Shortcuts by name and chooses menu bar commands.
Windows, scrolling, and keys
Left half, right half, maximize, minimize. It also scrolls and presses keys from a closed set.
AGI Cockpit
With AGI Cockpit installed, say "stop the current task" or "go with option B".
Automatic updates
Updates are delivered through Sparkle.
AGI Labo login
Sign in with an AGI Labo account that has an active membership.
Requirements
System
- macOS 26 (Tahoe) or later
- Apple Silicon
- An AGI Labo account with an active membership
Permissions
- Accessibility
- Microphone
- Speech Recognition
- Input Monitoring (only if the trigger does not respond)
Download
macOS 26 (Tahoe) or later / Apple Silicon / AGI Labo account with an active membership required