What Dictera sends over the network, and how to check
On default settings, the app itself talks to three places: our licensing service, the update feed on this site, and Hugging Face while a model downloads. None of them receives audio or text. Your words only leave the Mac if you point Dictera at a remote model yourself, and the app says so when you do.
What does the app send, and to whom?
Here is every request Dictera makes on its own, with the payload.
- The licensing service, on Google Cloud in the EU. It receives a SHA-256 hash of your Mac's hardware identifier and, once you have one, your license key. Activating a key also sends an activation name, which defaults to your Mac's name. It is checked at launch and then roughly once a day, at a random point between 20 and 28 hours later.
- dictera.app, for the update feed. The request carries nothing but itself; the reply is a list of versions. If an update is found, the download comes from the same site.
- Hugging Face, only while a speech model or the cleanup model downloads. You receive the weights; they see what any web server sees, such as your IP address.
Buying a license happens at Lemon Squeezy, in your browser, and they see what you type into the checkout. The security page keeps the same list as a table, and it is the version to forward to anyone who has to approve the app.
Why does the license need the network at all?
Subscriptions end on a server, not on your Mac. The daily check is how a lapsed subscription is noticed, and how the 14-day trial is kept to one per Mac even after a reinstall. That is why the hardware identifier is hashed before it leaves: the server needs to recognise the same Mac twice, not to know which Mac it is.
If the check cannot get through, nothing stops. The last good result is kept, and a license carries on offline for 14 days. Why there is no account covers what the licensing service holds and what it cannot.
What never goes anywhere?
Audio, to anyone, ever. It is transcribed as it arrives and never written to disk as a recording. Your transcripts, history, meeting notes and screenshots stay in a folder on your Mac. There is no analytics or crash-reporting library in the app, so there is nothing to switch off: it is not in the binary.
Both models that turn speech into writing run locally. Two models per dictation explains what each does; neither needs a connection once downloaded.
When does my text leave the Mac?
Only when you add a remote model. Dictera can send cleanup, translation or the assistant to a server of your choosing: one on your own Mac, one on your network, or a cloud provider with your own API key. Enter an address that is not on this Mac and the settings say, under the Test Connection button, that your text will leave your device. Over plain http, they add that the connection is unencrypted.
What goes is the text being processed. For the assistant, that is the transcript window and any screenshots you captured, sent on each pass. The Claude Code CLI engine is the same kind of choice: it sends your conversation, and files it reads in the folder you nominated, to Anthropic under your own account.
One thing does not follow your engine choice. To place text correctly, Dictera sometimes looks at the characters around your cursor. That fragment of your document only ever goes to the built-in model or a server on this Mac. With a cloud engine, the app decides by a local rule instead of sending it.
How can I check this myself?
Three ways, from quickest to most thorough:
- Look at the open connections. With Dictera running, run
lsof -i -P | grep -i dicterain Terminal, then dictate, translate and record a meeting. On default settings you will see nothing beyond an occasional license or update check. It shows open connections only, so run it more than once. - Turn the Wi-Fi off. Once the models are downloaded, dictation, cleanup, translation, correction and on-device meeting notes all keep working.
- Use a firewall. Little Snitch or LuLu show every connection as it is attempted, and let you refuse the ones you do not want.
Two things will not show up under Dictera's name. The Claude Code CLI runs as its own process, so its traffic appears under that. And Apple's speech engine gets its language files through macOS, the way it installs any system language asset.
Can I turn the scheduled requests off?
The update check, yes: Settings ▸ General ▸ Software Update has Automatically Check for Updates, and you can check by hand from the menu instead. The license check, no, but a firewall rule against it costs you nothing for 14 days, and letting it through once resets the clock.
Why we built it this way
A speech tool hears things that were never meant to be written down, and its privacy promise is only worth something if it can be checked. So we kept the list short, left out anything that phones home by default, and put a one-line Terminal command on the security page instead of reassurance. The privacy policy has the same list, with the legal bases.
