Hush

Dictation, file transcription and meeting notes in one macOS app. Local-first: speech runs on your Mac by default, and cloud speech or AI providers are opt-in.

• Dictation: hold fn anywhere and talk, then release to insert polished text.
  • A quick tap locks hands-free recording, and Esc cancels.
  • Command mode (fn⇧) rewrites selected text by voice.
  • Per-app styles, vocabulary, replacements and searchable history.
• Files: drop audio or video, or paste a YouTube or podcast link, to get a timestamped, searchable transcript.
  • Synced playback and inline editing.
  • Speaker names, translation, summaries and chat.
  • Export to Markdown, TXT, SRT, VTT, CSV or JSON.
• Meetings: a notepad-first recorder that captures your mic ("Me") and system audio ("Them") without a bot joining the call.
  • Live transcript.
  • Enhanced notes from templates, chat and recipes (follow-up email, action items, and more).
  • Detects when Zoom, Meet, Teams and similar apps start using the mic.

Meeting notes	File transcript
Dictation history and styles	AI providers

Install

1. Download macos-arm64-Hush.dmg from the latest release.

2. Open it and drag Hush to Applications.

3. Release builds are not notarized unless the repository has Apple signing secrets, so macOS may say Hush "is damaged" or "can't be opened". Clear the quarantine flag once:

   xattr -dr com.apple.quarantine /Applications/Hush.app

4. Launch Hush. Onboarding grants permissions and downloads a speech model (Parakeet v3, about 740 MB, is recommended).

Optional: brew install ffmpeg yt-dlp for more media formats and link imports.

Speech engines

On-device	transcribe.cpp on Metal (in the native helper), using Parakeet TDT v2/v3, Whisper (base → large v3 turbo), Moonshine, SenseVoice, Canary, Qwen3-ASR and others. Models download from Settings → Speech.
Cloud (BYOK)	OpenAI, Groq, Mistral, xAI Grok, Meta Muse, Deepgram, ElevenLabs, AssemblyAI, or any OpenAI-compatible /audio/transcriptions endpoint.

Dictation, Files and Meetings each choose their own engine.

AI providers

AI is used for dictation polish, command mode, meeting notes, summaries, translation, chat and recipes.

• Agent CLIs you're already signed in to: Claude Code, Codex, opencode, Gemini CLI, Antigravity.
  • They run with tools disabled or read-only, in an empty temporary directory. Claude Code, Codex and Gemini CLI keep no session history; opencode and Antigravity may keep their own.
• Bring your own key: any Chat Completions, Responses or Anthropic Messages endpoint.
  • Presets: OpenAI, Anthropic, OpenRouter, Groq, xAI, Ollama, LM Studio.

Development

Requirements: macOS 14.2 or later (for system audio capture), Bun, Xcode (Swift 6) and CMake (to build the native helper, which compiles transcribe.cpp). Optional: ffmpeg (more media formats) and yt-dlp (link imports).

Hush is an Electrobun app: the window is the bundled Chromium (CEF), the backend is TypeScript on Bun, and a small Swift helper (native/hush-native) does audio capture, speech-to-text, global hotkeys and paste.

bun install
bun run app:dev         # the app (builds the helper and UI, then electrobun dev)
bun run app:build       # release build: .app, .dmg and update files in artifacts/
bun run check           # svelte-check + tsc
bun run test            # bun tests for the TypeScript backend
bun run test:native      # swift test for the native helper

AGENTS.md has the code layout and conventions for contributors (human or AI).

CI and releases

• CI runs on every push and pull request: svelte-check, tsc, the bun tests and the UI build, then swift format lint --strict, a warnings-as-errors swift build and the unit tests of the native helper on macOS.

• Release runs when a v* tag is pushed. The tag must match version in electrobun.config.ts, package.json and helperVersion in native/hush-native/Sources/HushNative/Methods.swift (bun scripts/version.ts checks they agree). It builds the Apple Silicon .dmg with electrobun build and publishes a GitHub release. Tags with a - (e.g. v0.2.0-beta.1) become pre-releases.

  # bump version in electrobun.config.ts, package.json and helperVersion in Methods.swift, then:
  git tag v0.2.0 && git push origin v0.2.0

• To sign and notarize, add these repository secrets: ELECTROBUN_CERTIFICATE (base64 .p12), ELECTROBUN_CERTIFICATE_PASSWORD, ELECTROBUN_DEVELOPER_ID (the identity name), and for notarization ELECTROBUN_APPLEID, ELECTROBUN_APPLEIDPASS (app-specific password) and ELECTROBUN_TEAMID. Without them the build is unsigned. Locally, ELECTROBUN_DEVELOPER_ID="Hush Local Signing" bun run app:build signs with a keychain identity.

UI without the native side

bun run dev, then open http://127.0.0.1:1427 in a browser. When the page runs outside the app, src/lib/devmock fakes the IPC with fixture data.

• Add ?onboarding to the URL for the first-run flow, or ?live for a meeting in progress.
• window.__hushEmit(event, payload) fires any backend event. For example, open /overlay and send dictation:state and dictation:level.

Scripts

bun run transcribe -- talk.m4a [model-id]   # local STT on a file (builds the helper if needed)
bun run llm                                 # dictation cleanup via each installed CLI
bun run systap                              # system-audio tap smoke test (prints levels)

Set HUSH_DATA_DIR to run against a throwaway data folder instead of ~/Library/Application Support/com.hush.app.

Permissions

• Microphone: dictation and meetings.
• Accessibility: the global hotkey and pasting into other apps.
• System Audio Recording: the other side of meetings. macOS asks the first time a meeting records. It's under Privacy & Security → Screen & System Audio Recording.

Data lives in ~/Library/Application Support/com.hush.app: SQLite, audio, models, and settings.json (mode 0600, holds API keys). Electrobun also keeps the Chromium profile in com.hush.app/stable/.

The app is a new binary relative to older builds, so macOS may ask for these permissions again after upgrading.

See DESIGN.md for the research and architecture.

License

MIT