Dictation that never leaves your Mac
Just talk. It's typed in about a second.
Hold Right ⌥, speak, let go — clean text lands in whatever app you're using. No upload, no queue: transcription and cleanup run on your Mac, so your voice never leaves it.
Requires macOS 14+ · Apple SiliconFree & open source · v0.1.0
Built onwhisper.cpp+Apple MLX>15× real-time on Apple Silicon
What you said
um hey priya so the build is basically ready uh I just need to fix the signing step and run the tests one more time I think we can like ship it friday afternoon if nothing breaks
- To:
- Priya
- Subject:
- Friday release
What LocalFlow typed
Hey Priya, the build is ready. I just need to fix the signing step and run the tests one more time. I think we can ship it Friday afternoon if nothing breaks.
Pasted about a second after you let go
- No cloud
- No telemetry
- No accounts
- MIT licensed
Speed
No upload, no wait.
There's no upload and no server queue. The moment you release the key, your Mac transcribes, tidies and pastes — typically in about a second with the Base model.
You talk15 s
Transcription< 1 s
Three moves
How it works
One key, held. No app to switch to, no window to click — it works wherever your cursor is.
Hold Right ⌥
Anywhere — your editor, a chat window, a browser form. A small waveform shows LocalFlow is listening.
Speak naturally
Ums, restarts and missing punctuation are fine. Add your names and jargon to the vocabulary and they come out right.
Let go
Clean text is pasted where your cursor is, and your clipboard is put back exactly as it was.
Keyboard controls
Push-to-talk
Hold Right ⌥
Talk while the key is down. Let go to paste.
Hands-free
Double-tap Right ⌥, press once to stop
For long takes: lock it on, put your hands down, press once when you're done.
Cancel
Esc
Changed your mind? Esc stops it at any stage, and nothing is pasted.
Any app
Wherever your cursor is
If you can type there, you can talk there. LocalFlow pastes into the focused app — no plugins, no integrations, nothing to switch to.
Settings
Tuned to how you talk
Every card here is a real setting in LocalFlow. Choose how much it tidies, teach it your words, pick the model that fits your Mac.
Choose how much it tidies
Four cleanup levels, from your exact words to tidy lists. Light is the default.
- VocabularyPriyanameLocalFlowsounds like: local flowMLXsounds like: em el ex
Heard: ship the local flow build to priya
Typed: Ship the LocalFlow build to Priya.
Your words, spelled right
Add names and jargon once. They steer recognition, and “sounds like” aliases are fixed the same way every time.
Tiny · fast, rough78 MB
Base · recommended start148 MB
Small · best balance488 MB
Medium · high accuracy, slower1.5 GB
Large v3 Turbo · max accuracy1.6 GB
Pick your model
Start with Base. Bigger models trade disk space for accuracy, and each is downloaded from Hugging Face only when you ask.
Ship the LocalFlow build to Priya.
Recent transcripts
Settings…⌘,
Quit LocalFlow⌘Q
- Ship the LocalFlow build to Priya.
- Running ten minutes late — start without me and …
- Hey Priya, the build is ready. I just need to fi…
Copied — press ⌘V
Nothing you say gets lost
If a paste can't land, the text waits on your clipboard and the overlay tells you to press ⌘V. Your last ten transcripts sit in the menu bar, one click from your clipboard — kept in memory, never written to disk.
Privacy
Private by construction
Not a privacy setting — the architecture. There's no server for your voice to go to.
No cloud
Speech recognition and cleanup run on your Mac. Audio and text never leave it.
No telemetry
No analytics, no crash reporting, no accounts — not even optional ones.
One kind of network request
Downloading models from Hugging Face, and only when you ask. Everything else works offline.
Password fields are off-limits
LocalFlow detects secure fields and never types into them or writes to your clipboard there.
Under the hood
Five on-device stages, all inside the line. If any one fails, your words still land somewhere you can see them.
Hugging Face · Model downloads only, when you ask
Don't take our word for it
A privacy promise is only worth what you can check. Here are three ways to check this one.
Install
Three steps to your first sentence
Download, drag, allow two permissions. A setup window walks you through the rest, then you hold ⌥ and talk.
Then hold Right ⌥ anywhere and talk.
- Chip
- Apple M1 or later
- macOS
- Sonoma 14 or later
Find it: Apple menu → About This Mac
Intel Mac, or macOS 13 and earlier? LocalFlow won't run there. It is built for Apple Silicon only.
Requirements
- macOS 14 (Sonoma) or later
- LocalFlow lives in the menu bar; there's no Dock icon.
- Apple Silicon (M1 or later)
- Built for arm64 only — Intel Macs aren't supported.
- 148 MB to get started
- The recommended Base speech model. Speech models range from 78 MB to 1.6 GB; the optional cleanup LLM is 0.7 GB or 2.3 GB.
- Microphone and Accessibility access
- A setup window walks you through both permissions and the first model download.
FAQ
Fair questions
The honest answers, including the ones where the answer is no.
Is it really free?
Yes. LocalFlow is MIT licensed open source: no account, no subscription, no trial and no paid tier. You can read, build and change every line.
Does anything I say leave my Mac?
No. Audio is recorded, transcribed and cleaned up on your Mac, and the text is pasted locally. The only network traffic LocalFlow ever produces is downloading model files from Hugging Face, and only when you ask.
Does it work offline?
Yes, once a model is downloaded. You need an internet connection only while downloading models; everything else works offline.
Will it run on my Intel Mac?
No. LocalFlow is built for Apple Silicon (M1 or later) only, and needs macOS 14 (Sonoma) or later.
Which languages does it understand?
The speech models are multilingual, with automatic language detection. Larger models are more accurate; you can switch any time in Settings → Models.
Does it keep my transcripts?
Only your last ten, in memory, so you can copy one back from the menu bar. Nothing is written to disk, and quitting clears them.
What happens in password fields?
Nothing. LocalFlow detects secure fields and never types into them or writes to your clipboard there.
Hold ⌥ and start talking.
One download, one setup window, and every word stays on your Mac.
Free & open source · macOS 14+ · Apple Silicon · v0.1.0
- Signed & notarized
- SHA-256 published
- No account