---

Why I made it

This started as a hobby project. I wanted transcripts of my work calls, but I never liked recording my colleagues just to get them — I needed the words, not a sound file of everyone that outlives the meeting. I'd also gotten hooked on a push-to-talk dictation hotkey and didn't want to give it up. What nagged at me was keeping two separate tools for that: one for dictation, one for meetings. I wanted a single app that did both, kept everything on my machine, and left no recording behind.

So I wanted to see how far I could get building it myself. Hush is what I came up with.

---

What it is

• A push-to-talk dictation tool. Hold your hotkey (default Right ⌘ on macOS, Right Ctrl on Linux + Windows), speak, release. The transcript is on your clipboard, ready to paste, before you've moved your hands. No browser tab, no web service, no upload.
• A meeting transcription tool. Click Record in a call — or let Hush auto-start when a meeting app takes your mic (Zoom, Teams, Meet, Slack, and a dozen others; on by default, toggle in Settings). It captures your mic and the call's system audio in parallel, transcribes both locally with whisper.cpp, and labels them You / Remote — or Speaker 1, 2… with the optional local diarisation model. Parallel system audio and auto-detection are macOS-only for now (#106 / #107); Linux and Windows run meeting mode mic-only.
• Both, in one app, sharing one history. Most tools pick one lane. Hush does dictation and meetings, with one model load and one searchable history.

The audio never leaves your machine, and never lands on disk — it's processed in RAM and gone the moment the transcript is on your clipboard.

---

How it compares

"Free?" and "Open source?" are split on purpose — they're often conflated, but a free tier on closed-source SaaS is the vendor's to revoke, while an OSS license is yours to fork.

	Local?	Dictation?	Meetings?	Free?	Open source?	Platforms
Hush	✅	✅ hotkey	✅ mic + system audio in parallel (macOS)	✅	✅ Apache 2.0	macOS · Linux · Windows
OpenWhispr	✅	✅	✅ auto-detect Zoom / Teams	✅ local unlimited; cloud tier: 2k words/wk free, $8/mo Pro	✅ MIT	macOS · Linux · Windows
Whispering	✅	✅	—	✅	✅ AGPLv3	macOS · Linux · Windows
Buzz	✅	partial (live mic, no PTT into other apps)	file import only	✅	✅ MIT	macOS · Linux · Windows
Meetily	✅	—	✅ system audio, no bot	✅	✅ MIT	macOS · Linux · Windows
VoiceInk	✅	✅	—	✅	✅	macOS only
MacWhisper	✅	✅	partial (file import)	freemium	—	macOS only
Superwhisper	✅	✅	✅ free tier (Meeting Transcription)	freemium	—	macOS · Windows · iOS
Granola	cloud LLM	—	✅	freemium	—	macOS · Windows
Otter / Fireflies / Fathom	—	—	✅ (cloud bot or web)	freemium	—	web

The cross-platform OSS rows are the real competition. OpenWhispr is the closest — local dictation + meetings on the same three OSes — though its builds bundle an optional cloud tier (local stays unlimited; cloud caps at 2k words/week free). What sets Hush apart: dictation and parallel-source meeting capture in one app, one whisper.cpp load, one searchable history, with a privacy posture you can verify (below). The closest macOS app, VoiceInk, is what inspired Hush — see Acknowledgements.

---

Privacy: grep the source

Privacy is the point of Hush, so it's built to be verified rather than promised.

• Two outbound reqwest callers exist in the entire codebase. Both are user-initiated. Grep src-tauri/src for reqwest::Client if you don't believe it:
  • Whisper / speaker model downloads when you click Download in the model picker. HTTPS-only, host-pinned to huggingface.co / *.hf.co (one signed-CDN hop allowed for HF's storage backend), redirect-cap of 4, SHA-256 verified on every download. Once the model is cached, transcription is fully offline.
  • Manual update check when you click "Check for updates". Single read-only request to api.github.com/repos/khawkins98/Hush/releases/latest. No identifying headers beyond the default user agent.
• No telemetry. No analytics. No crash reporter. No startup beacon. Auto-update is not enabled by default.
• Audio never lands on disk. Mic + system-audio capture goes into audio_buffer::AudioRollingBuffer — a 30-second RAM ring. Nothing is written to a temp file, no WAV is staged. The transcript is the only persisted artefact.

If telemetry or auto-update ever ships, it'll be opt-in with a separate privacy review.

That's the posture that matters if you handle audio you can't upload — therapy, legal, journalism, research: nothing leaves your laptop, no account holds your call history, and you can audit or fork the source yourself.

---

Install

Hush isn't signed with an Apple Developer ID — the $99/year certificate isn't worth it for a solo hobby project — so macOS (and Windows) show a security warning on first launch. Clearing it is a one-time, ~30-second step.

macOS: Download the .dmg from Releases, open it, drag Hush.app to Applications. First launch shows a Gatekeeper warning — right-click → Open clears it once.

Linux / Windows: Releases → latest v* tag (.AppImage / .deb, or .msi / .exe). CI-built, not hands-on tested.

New to Hush? docs/getting-started.md is the full walkthrough — install, the one-time Gatekeeper / SmartScreen bypass, permissions, your first recording, and meeting capture. macOS permission troubleshooting (including the "was granted — now revoked" prompt after updates) lives in docs/macos-permissions.md.

Hush does not auto-update — check manually via Settings → About → Check for updates (or the Hush menu on macOS). Code-signing would remove the warnings; it's sponsorable at github.com/sponsors/khawkins98.

---

Platform support

macOS 26 is the primary target — I daily-drive it, and it's the only hands-on-tested platform. Linux and Windows build and test in CI but aren't hands-on tested (meeting mode is mic-only there; on Wayland, push-to-talk degrades to the toggle hotkey). Hands-on reports and platform fixes are very welcome — open an issue with repro steps and your version, or send a PR.

---

Engineering

For people who want to know how it's built before trusting it:

• A trait seam at every OS boundary — AudioCapture, Transcribe, Diarize, HistoryRepository, and friends are traits with hand-rolled mocks, so the IPC tests run with no real audio device, SQLite, or Tauri runtime.
• Event-driven meeting auto-detection — CoreAudio HAL listeners catch a mic going live, classify the frontmost app, and start or stop a session. No foreground polling.
• An append-only decision log — learnings.md is dated and captures the why behind every non-obvious call, including the parts that got reverted.

This isn't Electron with a mic icon: four native windows, each with its own capability file; a native menu bar; tray-icon templating that adapts to light/dark menu bars; and traffic-light permission health that tells "stale" from "revoked" in the macOS TCC store. Full stack, window topology, trait seams, and module map are in ARCHITECTURE.md.

---

Documentation

For	Read
Discovering Hush — what it does	This README + the live app (install it, open Settings → menus describe what each thing does)
Install + first recording	docs/getting-started.md — install, Gatekeeper / SmartScreen bypass, permissions, dictation, meetings. macOS TCC troubleshooting: docs/macos-permissions.md
What's shipped right now	STATUS.md (rolling snapshot), CHANGELOG.md (release-by-release record)
Attribution + legal posture	hush-prd.md §13.8 — black-box reimplementation discipline and VoiceInk attribution rationale (the rest of that spec is historical)
How it's built	ARCHITECTURE.md — stack, four-window topology, trait seams, meeting pump, module map
Running it locally / dev commands	docs/developing.md — setup, command reference, macOS quirks, test layers
Building + contributing	CONTRIBUTING.md, CLAUDE.md for the Claude-assisted contributor workflow
Why decisions were made	learnings.md — append-only engineering decision log
Reporting a vulnerability	SECURITY.md

---

Acknowledgements

Hush is a behavioural reimplementation of VoiceInk by Pax — a fantastic macOS-native dictation app that solved the local-whisper-with-good-UX problem first. Hush takes the same product concept, adds meeting capture as a peer feature, and ships cross-platform.

No VoiceInk source code was copied or referenced at any point during development. Design was derived from VoiceInk's public README and observable runtime behaviour — see hush-prd.md §13.8 for the full reasoning. If you like Hush, VoiceInk deserves a look too.

---

License

Apache 2.0 — see LICENSE.