The answer depends on what you turn on. SuprFlow transcribes speech on your Mac with a downloaded speech model. Its default text cleanup is local. If you choose a hosted AI provider for advanced formatting, SuprFlow sends the transcript text and formatting instructions to that provider. Release builds also contain usage telemetry, and the app contacts SuprFlow for account and license checks. Saying “nothing leaves your Mac” would hide those distinctions.
This page describes SuprFlow 1.0.0 and its settings, not a guarantee about every signed build or your Mac's network. If your work is confidential, use your organization's approved configuration and verify the installed release before dictating client material.
Follow one dictation through the app
- Recording and recognition. Press the configured hotkey and speak. The Mac microphone captures audio; WhisperKit performs speech recognition with a model downloaded to the Mac. The normal recording pipeline removes its temporary audio file after transcription. This is different from retaining the resulting words.
- Local cleanup. The default formatting setting performs punctuation, capitalization, filler removal, and other cleanup on the Mac. It does not require a hosted text model.
- Optional advanced formatting. If you select Cloud API, the transcript and a formatting prompt are sent directly to the selected provider or custom endpoint. The audio recording is not the API payload in the current release. Provider privacy, retention, and model-improvement terms apply to that transcript. A user-supplied key does not make the request local.
- Optional Ollama formatting. The default Ollama address is
localhost:11434. That is a local path when Ollama actually runs on the same Mac. The endpoint can be changed, so inspect it before treating Ollama as a guarantee of on-device processing. - History. SuprFlow keeps up to 100 recent transcription entries in an Application Support JSON file on the Mac. A history entry can contain the raw transcript, cleaned text, formatted text, mode/model details, and timestamp; it is not limited to the final wording. In SuprFlow 1.0.0, deleting one row may not persist after the app restarts; use Clear All and check again after relaunching. Clear All removes the app history file, but cannot remove copies in backups, pasted destinations, or other apps. Check your organization's retention policy.
- Other network traffic. The release build initializes TelemetryDeck and sends named usage events with model, duration, language, and text-length metadata. It also checks the SuprFlow account and license. That telemetry includes free-form fields: an editable model identifier and a hosted formatter's error response body can be forwarded, the latter as a truncated error string. A provider could echo transcript text in that response. The event schema also accepts a mode name; the app does not expose a custom-mode editor, so a user-authored name is a future risk if that feature becomes reachable. Do not treat current release telemetry as guaranteed content-free; test a signed build before relying on that claim.
- Destination context. At recording start, the app reads the active app, window title, browser URL, field label, and up to 300 characters of existing text in the focused field into local memory. In SuprFlow 1.0.0, the formatter request does not pass this context to the hosted provider. The app does use context to suggest a formatting mode. Organizations should review this Accessibility access as well as network transfer, and recheck it in each signed release.
- Insertion clipboard. The ordinary insertion path briefly puts the final text on the macOS clipboard, simulates Command-V, then restores the previous clipboard string after about half a second if nothing else changed it. SuprFlow's own clipboard-history feature is a separate setting, but another clipboard manager could observe that temporary value. Live-stream insertion takes a different path. Test the actual signed build and destination before promising that dictation never touches the clipboard.
- Local diagnostics. In SuprFlow 1.0.0, several transcription paths print recognized-word excerpts to console output, so treat console logs as potentially containing dictated words. Clear All does not address separately collected diagnostic or support logs. See the logging review and test the signed release with fictional text.
What about AI model training?
In the local recognition and local cleanup path, SuprFlow 1.0.0 does not submit the dictated words to a remote model-training service. That is a technical observation about this path. It is not a blanket contractual promise for every setting. Cloud API formatting sends transcript text to your selected provider; check the terms of the exact provider, plan, and endpoint before using it for confidential work. Provider retention, abuse monitoring, and training controls can differ. If your organization requires an approved no-training arrangement, the security or legal owner should verify it rather than rely on a marketing phrase.
A five-minute test before confidential use
- Open General and confirm the speech language and whether clipboard history is enabled.
- Open Formatting and confirm the selected provider. Use the default local cleanup for a local-only text path; if using Ollama, verify the endpoint is on the intended host. If using Cloud API, record the provider, endpoint, account, and approved data policy.
- Open Transcriptions and decide whether local history fits your retention policy. Do this with synthetic sample text, not a client record.
- Dictate a harmless sentence in the intended destination app. Check the inserted text, history entry, and whether any formatter changed a name, number, legal term, or quoted phrase.
- For an institutional assessment, test the signed build's network behavior with synthetic data and review telemetry, account, update, and model-download endpoints. Make an explicit decision about local device backups and clipboard history.
Questions people ask
Is SuprFlow fully offline?
Local speech recognition and default cleanup can run after the model is downloaded. Account/license checks, updates, telemetry in release builds, and any selected hosted formatter have their own network behavior. Offline dictation and “the app never makes a network request” are different claims.
Does SuprFlow upload my audio to a hosted formatting model?
The Cloud API formatting path sends the transcript text and instructions, not the audio recording. The speech-recognition path uses a local model. Verify the signed release and any other enabled integrations before making an organization-wide statement.
Is my transcript deleted after it appears in another app?
SuprFlow's history can retain recent transcripts locally. The destination app, clipboard, backups, and any chosen hosted formatter can have separate retention. Review each location before treating a history-clear action as complete deletion.
Can I use it for legal, banking, audit, or client information?
The local recognition path can reduce one category of transfer, but that alone does not approve a workflow. Check the formatter endpoint, telemetry, local history, clipboard, device security, recipient app, accuracy review, and your firm's rules. Ask the responsible compliance or legal owner for approval where required.