On-device by default,
transparent about the rest

How MetaWhisp Handles Your Data, by Mode

Privacy claims are only meaningful if they're true in every mode. Here's exactly how each part of MetaWhisp treats your data.

1. Default on-device transcription

When you dictate, MetaWhisp runs OpenAI's Whisper model on your Mac through Apple's Core ML framework, on the GPU and CPU. Your audio is processed locally, held in memory for the duration of transcription and then discarded, and nothing is sent to a server. This mode makes no network requests for transcription, which you can confirm with any network monitor or in airplane mode.

2. Optional cloud transcription

Cloud transcription sends your audio to a transcription provider for that request. On the Free plan it is off unless you add your own Groq or OpenAI key. With Pro, the app switches transcription to the cloud each time it starts, and the audio goes through MetaWhisp's server to Groq. MetaWhisp does not store your audio or transcripts on its servers. If you handle sensitive or regulated data, use on-device mode.

3. Optional AI post-processing

Structured mode, translation and the other AI features use the model you choose: a local model that keeps everything on your Mac, your own OpenAI or Cerebras key, or Pro. With a cloud route, the text is sent to that model, never your audio. The Screen Agent can also send one screenshot of an allowed window, but only after you give that its own separate permission. With no key, no Pro and no local model, these features stay off.

4. App updates

MetaWhisp automatically checks for updates so it can offer you new versions. This contacts the update server to compare your version against the latest release. It's a version check — no audio, transcripts, or usage data are sent.

5. One-time model download

On first use, MetaWhisp downloads a small starter model so you can dictate right away, then the full Whisper large-v3 model (about 3.2 GB, once). After that, on-device transcription works offline.

Where Audio Is Stored

In on-device mode, your audio exists only in memory during processing and is discarded when transcription completes. The one exception is for when something goes wrong: if a transcription fails, or you quit the app during a meeting, the recording is saved to a Recovery folder on your Mac so you can retry, and files there are deleted after 7 days. In cloud mode, audio is sent to the provider for that request and is not stored on MetaWhisp's servers.

No Telemetry, No Analytics

MetaWhisp contains no analytics SDKs and no crash-reporting services. We don't collect data about what you transcribe or how you use the app, and the Free plan needs no account. Pro signs you in, and its server counts the cloud minutes you use against the monthly limit. On Free with no key added, the only automatic network activity is the update check and the one-time model download.

Data Protection & GDPR

MetaWhisp is designed to minimize data processing. In the default on-device mode, no personal data leaves your device during transcription, so there is no server-side processing of your voice data by us. If you choose an optional cloud feature, the relevant provider processes that request; review their terms for regulated use. We do not sell or share personal data.

Open Architecture

MetaWhisp uses the open-source WhisperKit framework by Argmax for on-device inference, and MetaWhisp itself is open-source (github.com/metawhisp), so the privacy behavior described here can be audited in the code rather than just trusted.