Hold a key, say the thing, and your Mac does it.
Cue is an ambient, voice-first assistant that reads what you're working on and acts on it. Native Swift, speech recognised on-device, and a hotkey instead of yet another window.
Six things, and they all point at the same idea.
Voice-first, hotkey-driven
Option+Space, speak, let go. No app to switch to and no window to find. The HUD floats over whatever you were already doing and disappears when you are done.
Sees what you are doing
Cue reads the active app, the frontmost window, your selection and your clipboard, and can OCR the screen on demand. The answer feels like a colleague who was watching rather than a search box.
Actually takes action
Claude decides, the Accessibility API executes. Open apps, navigate, paste, press buttons, run Shortcuts, drive AppleScript. Assistants that only answer are chat boxes.
Native Mac, not a web view
Swift and SwiftUI top to bottom. The HUD is a non-activating panel, so it never steals focus from whatever you were typing into.
Local speech recognition
Whisper.cpp runs on-device with Metal acceleration. Your voice does not leave the Mac to be transcribed, and only the final text goes to Claude, only when you invite it.
Direct distribution, no sandbox
Signed and notarised DMG with Sparkle for updates. Cue needs Accessibility, global hotkeys and the microphone, and the App Store sandbox would gut all three.
Sensors, trigger, transcription, reasoning, action.
Each stage is cheap until the one before it earns the next. Nothing wakes Claude up for nothing.
- Passive sensorsActive app, selected text, clipboard. Cheap, local, always on. No screen capture, no mic, no model calls.
- Screen samplerScreenCaptureKit plus Vision OCR, sampled on demand or at a cadence you set. Off until you opt in.
- TriggerOption+Space for v1, a wake word for v2, then a small on-device classifier that quietly scores ambient moments as worth surfacing.
- TranscriptionWhisper.cpp on-device turns voice into text, with voice activity detection short-circuiting empty recordings.
- ReasoningContext, transcript and tool definitions go to Claude, and the streaming response starts rendering before the model has finished thinking.
- ActionClaude picks a tool and Cue executes it. Destructive actions ask first, everything is logged, and undo is one keystroke away.
The scope is deliberate, and staying out of these is part of it.
Not a ChatGPT clone
Chat is the fallback. The voice loop is the primary interface.
Not a code editor assistant
Cursor owns that and does it well. Cue lives in the OS, not the IDE.
Not a meeting transcriber
Granola owns that. Cue is for when you are working, not when you are in a call.
Not always-recording
Rewind owns that. Cue is invited in rather than ambient in the surveillance sense.
Not cross-platform
Mac first, Mac only, at least until the shape of the product is right.
Not heading to the App Store
Accessibility access cannot live inside App Store sandbox rules. That one is fundamental.
Cue is being built in public by Voidmode Studios. Founder seats are limited to 200 and priced to reward the people who back it before the product is finished. After that, pricing goes to £199 at launch.
Back it before it ships.
Founder seats are limited to 200 and priced at £99, against £199 at launch.