hush

Your computer should make room for your thoughts. hush lets you speak into the apps you already use, shape your words with local AI, and keep your flow. Speech recognition and AI inference run on your Mac.

Start with your voice

1. Download the notarized alpha, open the DMG, and drag hush into Applications.
2. Open hush and follow the guided microphone and Accessibility setup.
3. Put your cursor in a text field. Hold Fn, speak, and release to insert your words.

The bundled speech model works offline. Optional models download once through settings. Shortcuts are configurable; without Accessibility access, dictation can be copied from the clipboard and pasted manually.

Walk through your first dictation →

Requirements: an Apple silicon Mac running macOS 15 or later. We target 8 GB for transcription and recommend 16 GB for the assistant. Physical 8 GB validation is still pending. Intel Macs are not supported.

Make it your own

	In hush
Dictate	Speak where you work. Choose Whisper or Parakeet models and the language coverage you need.
Find your style	Polish, rewrite, or create your own Personas using Apple Intelligence where available or downloaded Qwen models.
Ask for a hand	The optional Assistant (Beta) uses local AI to understand your active window and carry out supported actions. Hold Control + Fn to talk; Escape stops the conversation.
Keep the story	Optional Verified Messages let recipients inspect a signed capture and edit record through a private link.

Choose a writing style or create your own. Model setup and removal stay beside the feature that uses them.

This is an alpha. The assistant is experimental, and desktop automation depends on what each app exposes. Setup, model choices, and behaviour may change. See the release notes for the published build.

Local by design

Dictation and local AI work without a cloud inference service or an account. The assistant requests screen access separately; ordinary dictation does not need Screen Recording permission.

Downloads and update checks use the network. If you enable Verified Messages, hush uploads an encrypted message record. The original audio is not uploaded; the decryption key stays in the private link, and recipients decrypt in their browser. A signed history does not establish identity, human presence, or truth.

Privacy · How Verified Messages work

Built for this Mac

Swift, AppKit, and SwiftUI provide the native interface. WhisperKit and FluidAudio handle speech; MLX runs the downloaded language and vision models. The assistant uses macOS Accessibility and ScreenCaptureKit, with Kokoro for local speech.

git clone https://github.com/Powermama/hush.git
cd hush
open macos/Hush/Hush.xcodeproj

Choose the Hush scheme and your signing team, then run with ⌘R. The project uses pinned Swift packages. See the development guide for toolchain requirements, testing, local proof-service setup, and packaging.

One project, two parts

Directory	What lives here
macos/	The native Swift app, local inference, desktop actions and distribution tools.
webapp/	The web workspace: the Vite website with browser verification and the Bun/Hono proof service (apps/web).

To run the website and service locally:

cd webapp
bun install --frozen-lockfile
bun run dev

Open http://localhost:5173. The API runs on port 3000. Neither service is needed for local dictation. The migration record explains the clean baseline and preservation of the earlier histories.

Build it with us

Good bug reports, accessibility improvements, documentation, and focused pull requests all help. Start with the contribution guide or open an issue.

Find your area in the journey map, or learn how the pieces fit together. We welcome clear explanations, thoughtful design feedback and people learning along the way.

If hush is useful to you, star the repository to help others discover it.

Star history

License and credits

hush uses GNU AGPLv3 only for its original code and documentation. Use it, remix it and share improvements under the same terms, including in covered modified hosted versions. Commercial use is allowed.

The earlier native MIT grant remains valid. Model weights, dependencies and artwork retain their own terms and attribution. Read licensing scope and contributions for the boundaries and release obligations.

Built with WhisperKit, FluidAudio, MLX Swift, MLX Swift LM, and Sparkle.