Open-source Wispr Flow clone for macOS — hold a key, speak, watch your words appear.
Private · Fast · Free · Open Source
- Hold your hotkey (default: Right ⌘)
- Speak
- Release — your words appear instantly in whatever you're typing
OpenFlow reads the surrounding screen text via macOS Accessibility APIs to give Whisper contextual hints, then pastes a draft immediately while an LLM silently cleans up filler words in the background. Total latency from release to final text: ~500ms.
Hold key → audio capture
│
├─ screen context (AX APIs) ──→ Whisper prompt bias
│
└─→ Whisper transcription → draft paste
│
LLM refinement (background)
│
replace draft in-place
Progressive paste: the raw transcription lands immediately. The LLM refinement runs in the background and silently replaces the text if it changed anything. You never wait.
| Wispr Flow | OpenFlow | |
|---|---|---|
| Price | $20/month | Free |
| Open source | No | Yes |
| Telemetry | Unknown | None |
| Custom LLM / model | No | Yes |
| Local model support | No | Planned |
| Latency | ~800ms | ~500ms |
| API key required | No (bundled) | Yes (Groq free tier) |
- macOS 12+
- Groq API key — free tier covers heavy daily usage
- Microphone, Accessibility, and Input Monitoring permissions
brew install vishnuhari17/tap/openflowDownload the latest .dmg from Releases, open it, and drag OpenFlow.app to the Applications folder. Then double-click OpenFlow.app from Applications to launch.
Important: Do not run the app directly from the DMG. You must drag it to
/Applicationsfirst, otherwise Launch at Login won't work and macOS may show a security prompt on every launch.
git clone https://github.com/vishnuhari17/OpenFlow
cd OpenFlow
cargo build --release
./target/release/openflow setup
./target/release/openflow live# 1. Add your Groq API key
echo 'GROQ_API_KEY=gsk_...' >> ~/.env
# 2. First-run wizard (checks permissions, writes default config)
openflow setup
# 3. Start
openflow liveOn first launch, macOS will prompt for Accessibility and Input Monitoring permissions. Both are required — OpenFlow uses them to read screen context and detect key presses.
Once the app is running, click the menubar icon and toggle "Launch at Login". This creates a LaunchAgent that starts OpenFlow automatically when you log in. Run openflow setup first to clear any quarantine attributes that would otherwise cause macOS to prompt on every login.
~/.config/openflow/config.toml (created automatically by openflow setup):
# Key to hold while speaking. Options: right_command, f18, f19
hold_key = "right_command"
# How much surrounding screen text to include in the Whisper prompt (chars).
max_screen_context_chars = 2000
# Optional: override the default Groq models.
# transcription_model = "whisper-large-v3-turbo"
# refinement_model = "llama-3.1-8b-instant"
# Optional: language hint for Whisper (ISO 639-1). Empty = auto-detect.
# language = "en"All fields are optional. Env vars (GROQ_TRANSCRIPTION_MODEL, GROQ_REFINEMENT_MODEL, GROQ_TRANSCRIPTION_LANGUAGE) override config-file values.
| Stage | Typical |
|---|---|
| Audio capture | ~200ms |
| Whisper via Groq | ~180ms |
| Draft paste | <10ms |
| LLM refinement (background) | ~120ms |
| Total to first visible text | ~390ms |
openflow setup # first-run wizard
openflow permissions # check permission status
openflow focus # inspect currently focused element (AX API)
openflow focus-after 3 # same, after 3s delay — switch to target app first
openflow ax-debug # dump full accessibility debug report
openflow paste "hello world" # test paste injection
openflow paste-after 3 "hello" # test paste injection after 3s delay
openflow monitor-hotkey # watch raw hotkey events (useful for finding key names)
openflow demo # dry run — no mic, no API callssrc/
├── main.rs Entry point; tray on main thread, pipeline on background thread
├── app.rs AssistantApp: command dispatch + live dictation loop
├── config.rs AppConfig: loads ~/.config/openflow/config.toml
├── tray.rs Menubar icon with state machine (idle/recording/processing/done)
├── pipeline.rs Generic Pipeline<A,F,P,R,T>: audio → transcript → paste
├── domain.rs Core types: AudioBuffer, ScreenContext, TranscriptDraft, etc.
├── services/
│ ├── audio.rs LiveAudioCapture: cpal + VAD + 16kHz downsample + WAV encode
│ ├── focus.rs MacOsFocusResolver: AX API → app name, window title, field preview
│ ├── paste.rs MacOsTextPaster: clipboard + ⌘V injection + replace_recent_paste
│ ├── transcription.rs GroqTranscriptionEngine: Whisper via Groq API
│ ├── refinement.rs GroqTranscriptRefiner: filler removal via LLaMA
│ ├── vocab.rs PersonalVocab: persistent per-user term frequency store
│ └── sound.rs Start/stop audio cues via afplay
└── platform/
└── macos.rs CGEventTap (hotkey), AXUIElement (accessibility), NSPasteboard
PRs welcome. The interesting unsolved problems:
- Local model mode — whisper.cpp via FFI or subprocess. Zero API calls, fully offline.
- Streaming LLM paste — word-by-word insertion as the LLM streams tokens.
- Auto-punctuation — detect sentence boundaries from pause duration.
- Language switching — detect language per-utterance rather than per-config.
- Settings UI — Tauri-based settings window to replace the config file.
MIT