Toolbay Flow is the first free tool in the Toolbay collection: open-source rebuilds of paid
software. It does what paid voice-dictation apps do, with speech recognition
(whisper.cpp) and text cleanup (llama.cpp
with an open Qwen model) running on your own hardware.
It's written from scratch (Tauri 2 + Rust + React). No code, assets or services from any commercial product are used.
Install
Windows 10/11: download Toolbay.Flow_x.y.z_x64-setup.exe from Releases
and double-click it. One install and it works: no admin prompt, nothing else to install, no models to pick. The installer
includes a small speech model, so you can dictate straight away, offline. Flow then picks the best models for your computer
by itself and fetches them quietly in the background (a fast GPU gets Large v3 Turbo plus a 4B cleanup model; an ordinary
laptop gets compact models that stay instant on the CPU).
macOS (Apple Silicon or Intel): download the .pkg installer from Releases (aarch64 for Apple Silicon, x64 for Intel). The macOS build is produced by CI but has
not been tested on a real Mac yet; see Status below. Grant Microphone, Accessibility and Input Monitoring when asked.
Until Toolbay Flow is notarized by Apple, opening the installer says Apple can't check it: click Done, open
System Settings → Privacy & Security, scroll down and click Open Anyway, then confirm. The installer puts
Toolbay Flow in Applications and it opens normally from then on.
Through the Toolbay app: open Toolbay, find Flow, click Install. Toolbay reads toolbay.json,
downloads the installer from the latest GitHub release and launches Flow in its own window.
How to use it
| Do this | Windows | macOS |
|---|
| Dictate (hold, speak, release) | Ctrl + Win | Fn |
| Hands-free (start/stop) | Ctrl + Win + Space, or double-tap Ctrl + Win | Fn + Space |
| Command mode (edit selected text by voice) | Ctrl + Win + Alt | Fn + Ctrl |
| Paste last dictation | Alt + Shift + Z | ⌘ + Ctrl + V |
| Copy last dictation | Alt + Shift + X | ⌘ + Ctrl + C |
| Cancel while listening | Esc | Esc |
| Transforms 1 / 2 / 3 on selected text | Win + Alt + 1/2/3 | Option + 1/2/3 |
Every shortcut can be changed in Shortcuts, including modifier-only combos.
Features
- Global push-to-talk and hands-free hotkeys with a low-level keyboard hook (modifier-only chords, double-tap to lock,
Esc to cancel, Start menu suppression for Win chords).
- Flow bar: a small always-on-top pill at the bottom of the screen with a live waveform, hands-free stop/cancel buttons,
processing and done states, error and "learned a word" toasts. It follows the monitor your mouse is on and is
click-through everywhere except the pill.
- Local speech recognition with whisper.cpp, models chosen automatically for your hardware (GPU, video memory,
RAM, cores), re-checked when hardware changes and stepped down if dictation gets slow. Manual choice under Advanced.
- Fast: the mic opens instantly with a short pre-roll, speech is transcribed in chunks while you talk, and only the
tail is left when you let go. On an RTX 4070 Ti, release-to-text is about 0.16 s for 3 s of speech, 0.3 s for 10 s
and 0.85 s for 30 s (p50, including AI cleanup).
- AI cleanup: removes fillers and stutters, applies spoken corrections ("6, no wait, 7" → "7"), punctuation, spoken
commands ("new line", "question mark"), lists, numbers and email layout. Runs locally through llama.cpp
(Qwen3 4B Instruct on GPU, Qwen3 1.7B for CPU-only machines), with a rule-based fallback.
Four levels: Off, Light (rules only), Standard, Polished.
- Bring your own key (optional): any OpenAI-compatible endpoint or Anthropic, if you prefer a cloud model.
- Styles per app category (personal messages, work messages, email, everything else): Formal, Casual, Very casual, Excited,
plus your own app → category rules.
- Context awareness: reads the focused app, window title, selected text and the text around the cursor (Windows UI Automation,
macOS Accessibility) so dictation continues your sentence naturally. Password fields are never read.
- Personal dictionary: words and names to always spell right (also fed to the speech model), replacement rules,
starring, usage counts, and auto-learning: fix a word after pasting and Flow adds it, with Undo.
- Snippets: say a cue ("my calendar link") and get the full text.
- Command mode: select text, hold the command key and say "make this friendlier" or "translate to Spanish";
with nothing selected, say "reply saying I'll be there at 5" to write new text.
- Transforms: three editable rewrite presets (Polish, Prompt engineer, Make it shorter) on hotkeys.
- Paste into any app via the clipboard and a simulated Ctrl/⌘+V, then your previous clipboard (all formats) is restored.
Dictations are kept out of Windows clipboard history.
Status
Full checklist: docs/FEATURES.md.
| Area | State |
|---|
| Windows 10/11 x64 | Built, installed and tested on Windows 11 with an RTX 4070 Ti |
| macOS | Configured and built in CI (Accessibility, CGEventTap hotkeys, Cmd+V paste, Metal engines). Not yet tested on a Mac. |
| Speech on AMD/Intel GPUs (Windows) | Runs on CPU for now (cleanup already uses Vulkan on any GPU) |
| Code signing | Prepared in CI; needs certificates (see docs/SIGNING.md). Until then Windows shows a SmartScreen prompt |
| Meeting notetaker, cloud sync, teams, mobile apps | Not included yet |
Contributing
Read AGENTS.md and CONTRIBUTING.md. The rules in short: simplicity first; one-step install
that includes the models; 1:1 feature and UX parity with Wispr Flow is the bar; everything free; local by default;
no telemetry; clean-room only; consumer-first.
Build from source
Requirements: Node 20+, Rust stable, and on Windows the Visual Studio C++ Build Tools.
npm install
npm run tauri dev # run in development
npm run tauri build # produce the installer in src-tauri/target/release/bundle
cargo test --manifest-path src-tauri/Cargo.toml
Code map:
| Path | What |
|---|
src-tauri/src/hotkeys/ | Chord matcher, Windows WH_KEYBOARD_LL hook, macOS listener |
src-tauri/src/audio.rs | Microphone capture (cpal), resampling to 16 kHz, levels |
src-tauri/src/engines.rs | Model/engine catalog, downloader, local whisper/llama servers |
src-tauri/src/llm.rs | Speech request, cleanup / command / transform prompts, BYO-key providers |
src-tauri/src/textproc.rs | Rule-based cleanup, dictionary, snippets, styles, learning |
src-tauri/src/context.rs | Focused app and text around the cursor (UI Automation / AX) |
src-tauri/src/paste.rs | Clipboard snapshot, paste, restore |
src-tauri/src/dictation.rs | The dictation state machine |
src/bar/ | The floating Flow bar |
src/hub/ | Main window: onboarding, history, insights, dictionary, snippets, style, transforms, scratchpad, models, shortcuts, settings |
Toolbay integration
toolbay.json describes the app for the Toolbay launcher: id, version, the release asset per platform, how to install it
silently (/S for the per-user NSIS installer) and how to launch it (--from-toolbay). Flow is single-instance: launching
it again focuses the open window.
Credits
License
MIT. See LICENSE.
