Handy

Why it's ranked here

Handy backs its pitch with real engineering. The Rust backend separates audio capture, voice detection, transcription, and system integration into distinct managers instead of one tangled process, and the project documents its own rough edges, a Bluetooth microphone quirk, a hardware-restricted key combo, Wayland paste failures, rather than hiding them. It ships through official installers plus community packages on every supported platform, and a command-line interface lets it run headless for scripting. That mix of real architecture and honesty about limits is what separates a working tool from a demo.

What's good

The model choice is genuinely useful rather than cosmetic: heavier Whisper variants run faster with a GPU, while the CPU-optimized alternative detects the spoken language automatically and stays usable on machines without one. Voice-activity detection trims silence before it reaches the model, so short holds don't waste a transcription pass. The command line goes beyond a toggle, it can batch-transcribe an audio file completely offline and emit the result as JSON, which makes the same engine usable from a script instead of only from the hotkey. A community-built Raycast extension shows the interface is open enough for others to build on.

Tradeoffs

The command-line surface is labeled a beta feature, so scripting against it today means accepting it may still change. On Linux the reliable path requires installing a separate text-input tool for your display server; skip that step and Handy falls back to a library with weaker Wayland support. A keyboard shortcut built around the fn or Globe key only fires on Apple's own keyboards, never on third-party ones, because of how that key is reported at the hardware level. And a Bluetooth headset microphone can audibly drop in quality mid-recording on macOS, a tradeoff of how Bluetooth handles simultaneous audio in and out.

How to use it well

This fits someone who wants private dictation into whatever app has focus, without paying for a cloud transcription subscription or worrying about audio leaving the machine. Pick the heavier model if a GPU is available for the best accuracy, or the lighter CPU-optimized one on a laptop without one. On Linux, install the display-server-specific text-input helper the project recommends rather than relying on the fallback. Wayland users who need a truly system-level shortcut should bind it through their desktop environment or window manager to the command-line toggle rather than the app's own hotkey handling. It is not the choice if you need shared or cloud-synced transcripts across devices.

Observed

Architecture

Desktop application built with Tauri, combining a Rust backend with a React and TypeScript frontend.

Packaging

Ships official installers for Windows, macOS, and Linux, alongside a Homebrew cask for macOS and a winget package for Windows.

Interfaces

Includes a command-line interface for remote-controlling a running instance (toggle recording, toggle post-processing, cancel) and for headless batch transcription of a WAV file, with JSON output.

Model support

Supports two local transcription model families: multiple Whisper size variants with GPU acceleration, and a CPU-optimized alternative with automatic language detection.

Voice detection

Uses Silero for voice activity detection to filter silence before transcription runs.

Linux support

On Linux, reliable text insertion depends on installing a display-server-specific helper (a tool for X11, a different one for Wayland); without one the app falls back to a library with reduced Wayland compatibility.

Testing

Repository includes Playwright end-to-end tests, a dedicated keyboard-handling test, and separate lint/format tooling for the frontend and the Rust backend.

Known limitations

README documents platform-specific limitations directly, including a Bluetooth microphone quality issue on macOS during recording, a hardware restriction on certain keyboard shortcuts, and Wayland focus/paste issues.