Hold a key, talk, get text — without the cloud listening

What it does

HEX is local-first voice dictation. Hold a shortcut (Option by default on macOS), speak, release, and the transcribed text lands at your cursor in whatever app you’re using. Transcription runs entirely on-device via a downloaded local speech model — no cloud service involved in ordinary dictation. It’s built in Rust, with the macOS app as the flagship and Windows and Linux betas alongside it.

The interesting bit

The privacy posture is unusually concrete: history keeps pasted text for seven days with entry and byte limits, never stores audio or full URLs, and every optional feature that does phone home (OpenCode-powered rewriting and Voice Action) is off by default and explicitly documented as sending data to your configured provider. Also notable: on Windows and Linux with a Vulkan driver, Parakeet streams transcription while you speak, so text arrives within about 100 ms of release.

Key highlights

• On-device transcription with swappable local models (Parakeet, Whisper) and per-model language/hint memory
• Modes system for corrections, app-specific behavior, and optional text transformations
• Experimental voice commands with custom TypeScript command files (~/.config/hex/hex.config.ts)
• Runs on macOS 15+ (Apple silicon), Windows 10/11 x64 and ARM64, and x86_64 Linux (X11 and wlroots Wayland)
• Signed builds with in-app updates; organizations can disable updates via a preference key

Caveats

• Windows builds aren’t code-signed yet, so SmartScreen will complain
• Website-aware modes and commands currently require Brave Browser
• Custom commands are executable local code with your permissions — the README is refreshingly blunt about trusting what you install

Verdict

If you dictate regularly and care where your audio goes, this is one of the more thoughtful options. If you live outside the supported platforms or want cloud-grade accuracy from a big provider’s model, look elsewhere.