A local-first voice input tool: hold a shortcut to speak, release, and Hex transcribes the audio to text on your device and pastes it into the current application; it also features an “Annotate” mode that lets you draw on the screen and speak while recording, sending keyframes, strokes, and transcriptions together to a coding agent.
Stars 2035 (+41 / 24h)
Reasons
- Standard dictation is performed entirely on-device, requiring no cloud transcription service, and does not save audio by default; you can switch between local models such as Parakeet and Whisper, each remembering its selected language; with Vulkan drivers, Parakeet outputs English text in real-time as you speak, with the author stating that text appears approximately 100 milliseconds after releasing the key
- Modes can correct names and idiomatic spellings before pasting, and can also rewrite text for specific applications; connecting to a running OpenCode instance allows for polishing, and annotations recorded via Annotate can be sent directly to an OpenCode session
- Includes a read-only command-line interface for querying dictation history and meeting notes, provides a TypeScript SDK, and offers official skills to teach agents how to write custom transformation rules
Usage
- macOS (Apple Silicon, macOS 15+): Download the signed DMG or run brew install --cask anomalyco/tap/hex, grant microphone and accessibility permissions, and download voice models in the setup wizard; by default, hold Option to speak, and press Option-Shift-S to start annotation
- Custom transformations: Run hex commands init to generate ~/.config/hex/hex.config.ts; let an agent help you write them: npx skills add anomalyco/hex --skill hex-personal-commands
Risks
- Windows and Linux are in beta: The Windows version lacks code signing and does not include meeting or local transcription interfaces; Linux only supports X11 (such as i3) and wlroots-based Wayland, while x86_64 requires a CPU supporting AVX2; downloading models requires access to huggingface.co
- Enabling OpenCode polishing sends transcription text, the foreground application name, and Brave website domains to your configured model provider; local diagnostic logs may contain transcription content, so check them before sharing; external PRs are not accepted; licensed under MIT