Hushflow

Wispr Flow-style dictation that runs entirely on your computer. · hushflow-coral.vercel.app

Hold a key, speak, let go: your words appear in whatever app you're typing in, with the fillers, false starts and self-corrections cleaned up. Speech recognition and cleanup both run locally, so nothing you say leaves your machine.

Hushflow is a fork of Handy by CJ Pais (MIT), built with Tauri (Rust + React).

Features

• Wispr-style hotkeys. Hold Fn (Mac) or Ctrl + Win (Windows) to talk; tap it instead to keep recording hands-free until you tap again. Esc cancels. Remappable.
• Works right away on a Mac. On macOS 26+ it starts with Apple's built-in speech recognizer (no download), fetches Parakeet in the background and switches to it once it's ready. You can also pick an engine yourself.
• Local cleanup ("Clean up my text"). A small on-device LLM (Gemma 4 E2B by default, or Qwen 3.5, or Apple Intelligence) removes filler words, applies "no wait, Tuesday"-style corrections, and formats lists and paragraphs in about 0.2 s. A rule-based pass covers you while the model downloads, and a guard pastes your words unchanged if the model ever answers instead of cleaning. The instructions are editable.
• Light on resources. The cleanup model sleeps after the idle time you choose and wakes while you're speaking.
• History, file transcription, a slim waveform overlay.

English only for now. See docs/fork/SPEC.md for the design decisions and benchmarks.

Download

Every push to main builds a Mac (Apple Silicon) app and a Windows installer: open the latest run under Actions → Main Branch Build and download the artifacts. Tagged versions are published under Releases.

The builds aren't code-signed yet:

• macOS: after moving Hushflow to Applications, run xattr -dr com.apple.quarantine /Applications/Hushflow.app (or right-click → Open) the first time.
• Windows: in the SmartScreen prompt, choose More info → Run anyway.

Build from source

Requires Rust and Bun. On macOS, Xcode 26 for the Apple speech and Apple Intelligence bridges.

bun install
bun scripts/fetch-llama.ts      # llama.cpp server for local cleanup
bun run tauri dev

bun run tauri build also signs update bundles, which needs the maintainer's key. Without it, build with bun run tauri build -c '{"bundle":{"createUpdaterArtifacts":false}}'.

Platform details are in BUILD.md; the upstream Handy README (troubleshooting, CLI flags, integrations) is kept at docs/UPSTREAM_README.md.

License

MIT, like Handy. See LICENSE.