mrkeyoor.com_
Tue 11 Aug 19:08 UTC
AI Toolsevaluationupdated 11 Aug 2026

Handy

Handy is a free desktop dictation app that turns speech into text locally, then pastes the result into whichever application you are using. It solves the privacy and subscription problems of cloud dictation by running Whisper or Parakeet models on Windows, macOS, and Linux without sending recordings away.

Verdict

Handy is the local dictation app to try first if privacy, price, and hackability matter more than uniform platform polish. Its basic speak-and-paste loop is excellent on a compatible machine, and the model choice is unusually approachable for non-specialists. macOS and Windows users should test microphone startup and paste behavior before relying on it, while Linux users should expect real integration work.

Setup3/5Easy desktop install, but models and OS permissions add work
Docs4/5Honest platform notes and unusually useful troubleshooting
Community5/5Large audience, frequent releases, and many outside contributors
Maturity3/5Daily-use core works, with platform and capture regressions open

Who it’s for

People who want private push-to-talk dictation across email, documents, chat, and code editors.
Developers and accessibility-minded users who want an MIT-licensed app they can inspect, modify, and extend.
Users with enough disk space and compute for local Whisper or Parakeet models, including CPU-only Parakeet use.
Multilingual speakers willing to compare models and tune shortcuts, microphones, and paste behavior for their system.

Who it’s NOT for

Linux users who expect identical behavior across every desktop: the README calls Wayland support limited, requires external typing tools, disables the overlay by default, and documents compositor-specific workarounds.
Anyone who needs guaranteed capture from the first syllable on Windows: a current version 0.9.5 report describes 0.5 to 3 seconds of speech being lost even on powerful hardware.
Managed Windows environments where any antivirus alert blocks installation: one current report says Defender flagged the 0.9.5 installer, so affected teams need security review rather than assuming the alert is harmless.
Users who cannot spare roughly 487 MB to 1.6 GB per bundled model or wait for a first model download.
People seeking polished support from a staffed vendor: a recent issue template says the project currently has one maintainer, despite substantial community contributions.

Setup reality

On macOS or Windows, installing the release and granting microphone plus accessibility permissions is close to the promised five-step start. You must still download a model, and the documented files range from roughly 487 MB to 1.6 GB. Linux needs more patience: install a typing utility suited to X11 or Wayland, possibly add a user to the input group, provide gtk-layer-shell, and configure the desktop's global shortcut. Restricted networks may require manual model placement with exact filenames and directory layouts. Building from source is a separate Tauri, Rust, and frontend toolchain job covered outside the quick start.

Dictation without a cloud account

Handy has a refreshingly narrow job: hold or press a shortcut, speak, and paste the transcription into the active application. Audio stays on the computer. There is no hosted account to create, per-minute bill to watch, or remote service receiving a draft email, medical note, password explanation, or private conversation. For users who dictate throughout the day, that local design is more than a philosophical benefit. It removes a recurring dependency from a basic input method.

The app filters silence with Silero voice activity detection, then offers Whisper-family models and Parakeet options. Whisper provides several size choices and can use available GPU acceleration. Parakeet V3 targets CPU use and automatically detects language. The README reports about five times real-time transcription on a mid-range i5, which is the project's own reference result rather than a promise for every machine. Model selection lets users trade storage, language needs, speed, and accuracy instead of accepting one opaque service model.

Handy pastes into the current text field, so it is not tied to a specific editor or meeting product. That system-wide behavior is its strongest practical advantage. It can cover messaging, issue trackers, documents, terminals, and any other application that accepts text. A transcript history and dictionary make repeated use more forgiving, while remote-control flags allow Raycast, a desktop shortcut, or another local tool to start and stop recording.

The ordinary install is simple, then models enter the picture

Most macOS and Windows users can download a release, install it, grant microphone and accessibility permissions, choose shortcuts, and start. Homebrew and winget packages also exist, though the README says those packages are not maintained by the Handy developers. The first meaningful delay is the model. Documented downloads range from a 487 MB Whisper Small file to a 1.6 GB Turbo file, with Parakeet choices around 478 MB to 731 MB. That is reasonable for local speech recognition, but not trivial on a small laptop or restricted connection.

The manual installation guide is excellent. It names each app-data location, expected filename, archive layout, and restart step. Handy can also discover custom Whisper GGML files and models already in the shared Hugging Face cache. This makes the app meaningfully forkable and adaptable, matching the project's stated goal better than a settings screen with a fixed vendor catalog would.

Building the application is a different proposition. Handy combines a React and TypeScript interface with a Rust backend, Tauri, cross-platform audio, native transcription libraries, keyboard hooks, and platform packaging. Contributors need the separate build guide and a target operating system suitable for the feature they are changing. Users should take the binary route unless they specifically want to modify the app.

Linux is supported, but it is not the easy path

The README is admirably candid about Linux. X11 commonly needs xdotool. Wayland needs wtype, dotool, or on Ubuntu 26.04 a configured ydotool. The app also links against gtk-layer-shell, and some systems need a renderer environment variable to launch reliably. Global shortcuts on Wayland may have to be created in GNOME, KDE, Sway, or Hyprland rather than captured directly by Handy.

The recording overlay is disabled by default on Linux because some compositors make it active and steal focus from the application that should receive the paste. A current KDE Plasma 6 report says the overlay disappears whenever a three-monitor Wayland setup is enabled, although transcription and paste still work. Another open Ubuntu X11 report describes occasional UI pauses after direct xdotool pasting. These are concrete desktop-integration costs, not faults in the speech models, and they are exactly the sort of friction that can make a dictation utility feel unreliable.

Current bugs touch the core interaction

Version 0.9.5 is an active bug-fix release with improvements across audio recovery, Windows pasting, model downloads, permissions, overlays, packaging, localization, and memory release. That pace is reassuring, but several current reports deserve attention. One Windows user reports that the first 0.5 to 3 seconds of recordings are frequently missing in 0.9.5, including with both Bluetooth and wired microphones. Because losing the opening words silently changes the meaning of dictated text, Windows users should pause briefly after the shortcut and test their exact microphone.

Another report shows models deleted outside Handy remaining marked as downloaded. The app can then sit at loading, accept the recording shortcut, and produce no text, while Delete and Rescan fail to clear the stale entry. A separate Fedora report found automatic acceleration selecting a Vulkan device slower than the CPU on that machine, with explicit CPU selection restoring its prior performance. These issues argue for keeping debug mode available and checking the selected model and accelerator when behavior changes after an update.

A Windows Defender report for the 0.9.5 installer is not proof of malware. It is still operationally important for workplaces where a detection ends the evaluation. Handy documents how to verify release artifacts with its Tauri updater public key and minisign. Security-sensitive users should verify the downloaded artifact and follow organizational policy, not disable protection based on an assumption that the alert is false.

A thriving young project, not a finished appliance

Handy had more than 29,000 stars and 143 open issues and pull requests on August 11, 2026. The repository was pushed that day, only three days after release 0.9.5. That release credited many first-time contributors, evidence that the community is doing real implementation work rather than only starring the repository. The issue template also says there is currently one maintainer, which matters when a cross-platform application attracts this much demand.

The project is young and its settings system, shortcut handling, and logging are still being reworked. Still, its documentation, signed releases, active fixes, and transparent limitations inspire more confidence than a glossy app that hides local-inference tradeoffs. Try Handy if you want private everyday dictation and can spend an hour testing models, permissions, and paste behavior. Keep an alternative input method ready until it proves dependable on your particular desktop.

Alternatives

ProjectWhat it isPick it when
BuzzA cross-platform desktop transcription and translation app built around Whisper.pick this instead when transcribing or translating existing audio and video files matters more than system-wide push-to-talk dictation.
nerd-dictationAn offline speech-to-text utility for Linux desktops with a scriptable command-line workflow.pick this instead when Linux is your only target and you prefer a small, keyboard-driven utility over a Tauri settings app.
whisper.cppA portable C and C++ implementation of Whisper for local inference and custom integrations.pick this instead when you need an inference building block or command-line transcription engine and will build the desktop workflow yourself.

What people are saying

  1. [github-trending] cjpais/Handy

Sources

  1. Handy README
  2. Handy v0.9.5 release
  3. Windows opening-audio loss report
  4. Wayland multi-monitor overlay report
  5. Stale model state report
  6. Windows Defender installer report