Privacy model
What leaves your Mac, and when.
Zerm listens to your microphone and can type into any application. Two capabilities worth being specific about. This page says exactly what happens to your audio, your text, and your permissions — and none of it requires taking our word for it, because the source is public.
The short version
On the default configuration, nothing leaves your Mac. Speech-to-text, text-to-speech, and the enhancement model all run locally. There is no account, no telemetry service, and no server belonging to us in the path. Cloud providers exist as an option, and they stay dormant until you add a key and select one.
Audio
- Recording happens through the standard macOS microphone APIs, which means the orange recording indicator in the menu bar applies to Zerm exactly as it does to any other app.
- On the local transcription path — Parakeet, Whisper, or Apple Speech — the audio is processed on the machine and is never transmitted. That includes files you transcribe, and identifying their speakers.
- If you select a cloud transcription provider, the audio for that request is sent to that provider under your own API key, subject to their terms. This only happens for providers you configure yourself.
- Recordings are kept locally alongside their transcripts in History, and can be deleted individually or wholesale, or automatically after a period you choose.
- Zerm records your microphone only. It does not capture system audio.
Text
- Transcripts are stored locally in History so you can search, copy, or retry them. Deleting an entry deletes it.
- AI enhancement runs only when your Output setting uses it. The on-device provider keeps it on your Mac; choosing a cloud provider sends the transcript to that provider instead.
- Storage of AI request payloads is a separate setting from making the requests at all. Turning on enhancement does not silently start logging what you dictated.
- Read Aloud reads the current selection. Smart Reading cleans that text before it is spoken — on-device unless you have picked a cloud voice.
Permissions, and what each one buys
| Permission | Needed for | If you decline |
|---|---|---|
| Microphone | Recording anything at all | Dictation cannot work |
| Accessibility | Pasting at your cursor, reading the current selection | Zerm falls back to the clipboard |
| Screen Recording | Screen context for enhancement, in Enhanced output | Enhancement works from the transcript alone; everything else works |
| Automation (per browser) | Power Mode reading the current browser URL | Website triggers never match; app triggers still work |
Each is requested in context, at the point the feature needs it, rather than in a wall of prompts at first launch. Setup failures are surfaced instead of being swallowed.
Power Mode context
Power Mode adapts dictation to where you are writing, which means it reads the frontmost application and, for browsers, the current URL. That context is used only to pick the Power Mode for the recording. It is not accumulated into a history or a profile.
Network
Zerm reaches the network for four things: downloading the models on first use, checking for application updates, fetching announcements from this site, and any cloud provider you have explicitly configured. With models already downloaded and no cloud provider selected, the dictation and read-aloud loop works with the network off entirely.
Verifying any of this
Zerm is GPLv3. The source for every claim on this page is public, and the parts that matter most — the recorder, the transcription pipeline, the paste path, and the provider clients — are readable without building anything.