Where your audio goes
Last updated 17 November 2026. Every claim on this page can be checked against the app, because it describes what the code does rather than what we would like it to do.
Speech
Recognition runs on this Mac, or in the cloud when Cloud is on. On Local, audio is decoded in memory and dropped once the text is inserted, with two exceptions, both of them files you can open. Turning on keep dictations keeps the last twenty recordings in ~/Library/Application Support/Inmost/takes/ (and never more than 200 MB), so a wrong transcript can be replayed through another engine instead of argued about. A cloud dictation that fails before any text came back is kept in ~/Library/Application Support/Inmost/kept-takes/, ten at most, rather than lost while you decide whether to retry. Clear history deletes both, and switching keep dictations off deletes the first. There's no account audio is filed under.
There is one switch that sends a recording off this Mac, and a new Mac starts with it on. Cloud dictation sends that one recording to Microsoft AI through cloud.inmost.app. If Microsoft fails or takes too long, the same recording also goes to xAI. Each gets the recording, the language you chose and your dictionary words. Microsoft says it keeps nothing. xAI says it keeps a copy for 30 days. Turning it on from Settings asks first. Turning it off stops recordings leaving this Mac at once, and the local engine is always still there.
A second switch sends text and never a recording. Cloud rewrite is off until you turn it on and confirm. It sends the words of a dictation, and your dictionary words, to cloud.inmost.app, then to Baseten. If Baseten fails or takes too long, they also go to Fireworks. Both say they keep no prompt and no answer. We keep a count of rewrites and what each cost us, never the words. The two switches are separate: turning one off does not turn the other off.
If you choose a local model that runs behind a server, that server is on 127.0.0.1, your own machine. Whisper Magic talks to it over the loopback interface, which doesn't touch a network.
Transcripts, logs and history
Transcripts are stored as plain JSON lines in ~/Library/Application Support/Inmost/history/. You choose how long they are kept in Settings (7, 30 or 90 days, or forever), and anything older is deleted while the app runs, not only when it starts. The same window covers the fixes learned from your corrections and the diagnostics lines that describe each dictation; a rule you type yourself in Settings → Rules is yours and stays. Clear History deletes the transcripts, those learned fixes and those lines.
Learn from the words I fix is on unless you turn it off. While it is on, the words you change after a dictation are kept beside those transcripts, in ~/Library/Application Support/Inmost/history/corrections/: the phrase the transcript got and the phrase you left behind, the app it landed in, and the time. Nothing else is kept: no audio, no window title, no document. Turning it off stops new rows and deletes nothing; DELETE WHAT IT LEARNED removes what it has, and Clear History and the retention setting delete these rows exactly as they delete transcripts, and the fix each row produced, because the row is the only record of where that fix came from. A model trained from them would live in ~/Library/Application Support/Inmost/adapters/ and is deleted by the same actions.
Two files describe each dictation, both in ~/Library/Application Support/Inmost/logs/. The diagnostics record holds audio duration, loudness, how many partial transcripts arrived and how the utterance ended, and it ages out on the same window as the transcripts. It holds no transcript text. The log is prose for debugging and can contain the last partial transcript the pill showed, because a pill that displays the wrong words is a bug worth reading back. The log rotates at 1 MB and keeps five files.
Settings are stored in the app's preferences. Downloaded model weights live in the model server's own cache, not in the app's folder.
Diagnostics
Diagnostics exist so a bug can be diagnosed from evidence instead of guesswork, and they leave the Mac only inside a feedback report you send yourself. Crash reports are the one exception, and only with that switch on: the next launch sends the report a crash wrote.
Crash reports
Crash reports are off until you turn them on in Settings, and nothing asks you to. With the switch on, a crash writes a small report on your Mac and sends it to the same Worker: the app and macOS version, your Mac's architecture, the exception or signal name, the top of the stack as symbol names, and a random install id. It never carries what you dictated, the words you added to the dictionary, any setting, or a file path, and Settings shows the exact bytes before you switch it on.
Feedback
Sending feedback is an explicit action. The window shows the exact bytes that will be uploaded before you press Send, rendered from the same value that is posted.
Always in the report: your message, the kind of report, and a random install id minted on first launch. That id is not derived from hardware, an account or an email, so it cannot join you across devices. Your email address goes in only if you type one.
Three attachments, each its own switch:
- The last 64 KB of the log, on by default. It can contain recent partial transcript text.
- Your most recent dictation, off by default.
- App version, macOS version and architecture, on by default.
Reports go to feedback.inmost.app, a Cloudflare Worker with a single database, and are read by the person who writes the app. The Worker also records the network each report arrives from (an IPv4 address, or the IPv6 /64 your Mac's addresses sit in), because that is what stops one sender from filling the queue; it is stored with the report and deleted with it. If that address doesn't resolve, the app retries the same worker on its workers.dev address once, then keeps the report in a local outbox until one of them answers. Reports stay in that database until we delete them. Write to privacy@inmost.app and we delete yours.
License
The license is a signed statement, not a lookup. The app verifies the signature on your Mac with a public key built into it, so dictation never asks a server for permission and works on a plane with the Wi-Fi off.
There is one call, once, when a Mac is set up: the license and a short hash of your Mac's hardware identifier go to license.inmost.app, so one license can't quietly become ten. That hash is sixteen characters of a SHA-256, not the identifier itself, and it can't be reversed into one. It is stored against the license and nothing else. Reinstalling computes the same hash, so it doesn't read as a new machine. If the call fails, for no network or a server having a day, the Mac is licensed anyway. Nothing about your audio, transcripts, usage or settings is in it.
Update checks
Whisper Magic checks for updates once a day by fetching a static file, https://inmost.app/appcast.xml. The request carries the app version, the macOS version and the CPU architecture, plus the IP address any web request carries. It carries no other identifier, no audio, no transcript and no usage. Automatic checks can be turned off in Settings, and dictation never depends on them. Updates are handled by Sparkle, a third-party updater that fetches that file and checks a signature, and does nothing else.
What Whisper Magic doesn't do
- No analytics, no telemetry, no crash report that carries speech.
- No advertising identifiers, no third-party analytics or advertising code.
- No selling or sharing of data, because none is collected to share.
What would break this page
Adding a network call that carries audio or text, adding an analytics library, or turning cloud dictation on by default. Each one is a code change, and this page changes in the same commit as the code. If you find a claim here the app does not keep, that is a bug: write to privacy@inmost.app.
Android
The Android app has its own policy, because it keeps different things on the phone: inmost.app/android/privacy.
Questions
Write to privacy@inmost.app, or use the contact form. A person answers.