Don't trust us.
Verify it yourself.
Most privacy pages are vibes. This one is grep commands. Every claim below is something you can check in less than sixty seconds.
#The short version
WhisperDictation runs entirely on your Mac. Your audio never leaves your machine. There is no server. There is no telemetry. There is no account. When we say "100% local," we mean it literally — and the source code is right there for you to confirm.
#What we don't do
- ✗ Send your audio anywhere — it's processed locally and discarded
- ✗ Send your transcriptions anywhere — they go straight to your cursor
- ✗ Track you — no analytics, no telemetry, no crash reports
- ✗ Phone home — no update server, no license check, no pings
- ✗ Read the clipboard — we type at the cursor, we don't touch paste
- ✗ Load web fonts — the app is 100% native, no browser involved
- ✗ Use any external service — except HuggingFace, when you click Download Model
#Don't take our word for it
The entire app is ~3,400 lines of Swift. You can audit every network-facing
line of code in a single grep. Try this:
WhisperDictation/Engine/ModelManager.swift: URLSession.shared.download(from: model.url, ...)
Read ModelManager.swift → and see for yourself. Models come from HuggingFace because that's where the Whisper model files live — downloaded once, stored locally, never touched again.
#Permission receipt
Here's exactly what the app does with the permissions it asks for:
Microphone
- Captures audio only while you hold the hotkey
- Buffer stays in RAM
- Converted to 16kHz mono, fed to Whisper
- Discarded after transcription
- Never written to disk
Accessibility
- Listens for your hotkey (CGEvent tap)
- Types transcribed text at cursor
- Does not read screen contents
- Does not read other keystrokes
- Does not record what you type
#Build it yourself
Paranoid? We respect that. You never have to trust the DMG on our releases page.
Clone the repo, run make whisper && make app,
and you'll have a binary built from source you can read line by line.
Want to be really sure? Run Little Snitch or LuLu while you dictate. You'll see zero outbound connections. (Except when you click the Download Model button. Then you'll see one. To HuggingFace. We told you.)
#Reporting security issues
Found something concerning? Open an issue on GitHub or see our SECURITY.md for responsible disclosure.