The whole policy in one line: our servers know who you are. They never know what you said. Your recordings and your words stay on your Mac; during recognition the audio goes straight from your Mac to the provider you picked — there is no stop called “our cloud” in the middle. Below, for every kind of data: where it goes, where it sits, how to delete it — and for each one, how to verify it with your own hands.
You speak — the audio leaves from your Mac and nowhere else
During recognition the audio goes straight to the one provider behind the tier you picked — which provider sits behind which tier is stated live in the app’s own “Privacy” panel (it is updated as the tiers change, so it is newer than this page). Pick on-device recognition and not one byte leaves this computer. Recording files sit on your Mac, kept 30 days by default so you can re-run them, then deleted automatically.
Settings can cut that to 7 days, or to “don’t keep them at all”. Delete the local folder and they are gone for real. Want it fully offline? Pick on-device recognition.
With AI cleanup, translation or ask-anything switched on, the transcript is handed to a language model once (which one, and whether it passes through our servers, is stated live in the app’s own “Privacy” panel); the finished text and your history live only in a database on your Mac, synced to no cloud at all.
Switch cleanup off (the “raw transcript, untouched” setting) and the text reaches no language model. History can be deleted one entry at a time or wiped entirely, and you set how long it is kept.
So it hears the names and terms you say often, the words in your dictionary are sent along with every recognition request. That is written in plain sight in Settings, not buried in terms.
Delete an entry whenever you like; once deleted it stops being sent.
At the moment you start speaking it reads the visible text of the frontmost window once, to get proper nouns right. It lives in memory for that one request: never written to the database, never to disk, never to a log — and words from your screen never enter your dictionary.
One switch in Settings → Recognition turns it off. With on-device dictation it is never read in the first place.
API keys sit in the macOS Keychain (system-level encryption), settings in system preferences — both on this machine only. We cannot see them and never receive them.
Deleting the app cleans them up the way macOS normally does; keys can be cleared in Settings at any time.
We never handle your audio or your words. They go from your Mac straight to the provider you picked, through none of our servers.
No behavioural tracking. There is no analytics SDK in the app. Which app you were speaking into, and for how long, is never reported back to us.
No selling your data, no sharing it. There are no advertisers and no “partners”. Your data is not the product.
This website doesn’t track you either. There is no third-party analytics script on this page — open your browser’s network panel right now and check.
~/Library/Application Support/Murmur/ — the recordings and every word you have ever said to it are in there. Open it whenever you want. Delete it and it is really deleted.
Point a free firewall like LuLu at Murmur and watch its outbound connections: you will see the domains of the recognition and language-model providers you picked, and nothing else.
“Where did this go?” on the app’s History page reports it honestly, against your current settings: which recognition tier handled this entry, whether the text passed through a language model, whether screen context was on.
When accounts and subscriptions ship, our servers will handle exactly one thing — who you are: sign-in and subscription status, and that is the end of the list. Not one word of your recordings, transcripts, history or dictionary will be uploaded — that sentence is on this page in black and white, so on launch day you can hold the product up against it. Every change to this page leaves its date at the bottom.