For IT and procurement
Local-only configuration for government and regulated teams.
Blurt turns speech into text on the computer. Audio never leaves the device. This page lists exactly what crosses the network, and how IT locks the app into Private mode so nothing dictated ever can.
Last updated: August 22, 2026
How dictation works
You hold a key and talk. The microphone audio goes to a speech engine running on the same computer (on macOS 26, Apple's on-device speech framework; on macOS 14 to 25, WhisperKit; on Windows, whisper.cpp). The text is typed at your cursor. The audio is discarded. There is no server in that loop, and there is no way to turn one on.
The one optional cloud feature is AI cleanup, a formatting pass on the dictated text. It is off by default, it is a paid feature, and Private mode makes it unreachable.
What crosses the network, and when
| Traffic | What is sent | When | Blocked by Private mode? |
|---|---|---|---|
| Sign-in | Email address and a one-time code (Supabase, US) | Sign-in and token refresh | No. Sign-in is required once. |
| Plan check | Account token only | App launch, Home window | No |
| Word counts | A number: words in the dictation just finished. Never the words. | After each dictation | No. Counts apply the plan limit. |
| AI cleanup | The dictated text (only when turned on, only on Premium) | After a dictation, if on | Yes. Unreachable. |
| Update check | A request to the public release feed on GitHub (Mac: Sparkle, daily; Windows: Velopack, at launch) | Launch and daily | No. Mac: IT can turn it off with the Sparkle managed preference SUEnableAutomaticChecks. |
| Speech model download | One-time download of the on-device model (Mac 14 to 25 and Windows: from huggingface.co; macOS 26: Apple's speech assets) | First run, or after changing the model | No. Pre-stage the model for machines without internet. |
| Feedback | Only what a user types into Report an Issue | When a user sends it | No (user-initiated) |
Nothing else. There is no analytics, no crash reporting, and no telemetry in any Blurt client.
What Private mode locks
- Cloud AI cleanup is off and unreachable. The check happens where the network call would be made, not only in the settings screen.
- Transcripts are never written to disk. Recent dictations stay in memory until the app quits. Any existing history file is deleted.
- On Windows, the log file records only counts, never text.
- The menu bar and the Home window show an on-device indicator so the state is always visible.
With Private mode on, the only account traffic is sign-in, the plan check, and word counts (numbers). Dictated text and audio never leave the computer.
Lock it on with a policy (IT)
One managed value per platform forces Private mode on regardless of the user's setting or plan. The switch shows on and cannot be turned off. Read at launch.
Mac
Preference domain com.sds.blurt, key ForcePrivateMode (boolean). Deploy a configuration profile with a managed preferences payload for that domain (Jamf, Kandji, Mosyle, Intune), or run once as an administrator:
sudo defaults write /Library/Preferences/com.sds.blurt ForcePrivateMode -bool trueOptional: add SUEnableAutomaticChecks = false to the same profile to stop automatic update checks (Sparkle honors managed preferences).
Windows
Registry value ForcePrivateMode (DWORD 1) under HKLM\SOFTWARE\Policies\Blurt. Group Policy Preferences, Intune, or:
reg add HKLM\SOFTWARE\Policies\Blurt /v ForcePrivateMode /t REG_DWORD /d 1 /fSilent install: Blurt-win-Setup.exe --silent (per user). Mac: copy Blurt.app to /Applications. Full IT reference with verification steps: email support@justblurt.com for the managed-preferences guide.
Machines with no internet
Sign in once and let the first run download the speech model while the machine is online, or pre-stage the model file. After that, dictation works with no network at all. Word counts queue locally. Update checks fail quietly.
What stays on the device, and what we hold
| Data | Where | Notes |
|---|---|---|
| Microphone audio | Memory only, on the device | Discarded after transcription. Never written to disk, never sent. |
| Transcript history | Device only (off in Private mode) | Mac: Application Support. Windows: encrypted with DPAPI. Clearable by the user. |
| Settings and word-count totals | Device | No dictation content. |
| Account record | Blurt's servers (Supabase and Vercel, United States) | Email address, plan, weekly word totals. No audio, no text. |
| Payment details | Stripe | Blurt never sees card numbers. |
Statements for your review
FedRAMP, TX-RAMP, CJIS. In Private mode no dictated audio or text is processed on any server, so cloud authorizations for the dictation path do not apply; there is no cloud dictation path. The account service (email, plan, word counts) runs on Vercel and Supabase in the United States and holds no dictated content. For criminal-justice use, dictated text goes straight from the device into your records system; Blurt never stores or transmits it.
Student data. Blurt signs the Texas student data privacy agreement (TX-NDPA) and equivalent agreements. Our privacy policy names our legal entity, every subprocessor, retention periods, and the 60-day deletion commitment.
Accessibility. The accessibility conformance report (VPAT 2.5, Section 508 edition) is published and updated with each release.
This page describes the product's behavior. It is not legal advice, and it does not by itself make any deployment compliant with a given program; your own policies and agreements do that.