SuprFlow's Cloud API formatter sends text, not the microphone recording. SuprFlow first recognizes speech with a model on the Mac. Its local cleanup pass then edits that transcript. When Cloud API formatting is selected, the app sends the pre-cleaned words between transcription markers, a formatting instruction, a model name, and generation settings in a JSON request to the selected provider. The request authenticates with the API key the user configured.
That distinction reduces one kind of exposure, but the text may be just as confidential as the recording. A dictated client name, account fact, legal instruction, or audit exception becomes content in the provider request. Calling the first step “local” must not imply that the optional second step is local.
Follow the optional request
- The Mac captures audio and WhisperKit creates a transcript using a downloaded model. The ordinary recording pipeline removes the temporary audio file after transcription in SuprFlow 1.0.0.
- SuprFlow runs a local cleanup pass before the formatter. This can alter punctuation or words; compare the raw and cleaned versions if exact wording matters.
- The Cloud API path wraps the pre-cleaned text between
---BEGIN TRANSCRIPTION---and---END TRANSCRIPTION---. The chosen formatting prompt is sent alongside it. Depending on provider, the JSON uses Anthropic's message format or an OpenAI-compatible chat format. - The request goes to a built-in provider endpoint or a user-entered custom URL. A custom endpoint is a destination chosen by configuration, not automatically a SuprFlow-operated service.
- The response is post-processed and may be rejected by guardrails; on failure the app falls back to local cleanup. A failed API call can still have sent the transcript to the endpoint before the fallback appeared.
SuprFlow 1.0.0 lists providers including OpenAI, Anthropic, Gemini, xAI, DeepSeek, Mistral, Groq, and OpenRouter, plus a custom endpoint. Provider lists and terms can change. Record the exact endpoint, model, account plan, and contract used in your firm; a generic “no AI training” promise is not enough to establish retention and allowed processing for that route.
A test a security reviewer can repeat
Use a fabricated paragraph such as: “Client Cedar did not authorize the $14,200 transfer on 7 May.” Choose Cloud API on a test account approved for this exercise. Record the signed SuprFlow version and the selected provider. In a controlled test build or permitted network inspection, verify that the body contains the fabricated transcript and prompt but no audio file. Check the provider's request logs and retention terms where available. Then force a harmless formatter error: in SuprFlow 1.0.0, a non-200 response body can enter an analytics error string. That failure path needs its own test before anyone claims telemetry is content-free.
Do not use real client data merely to test a provider. A synthetic example is enough to establish the data path. Do not intercept encrypted traffic by weakening trust settings on a production Mac; use a test environment and an approved method.
Questions people ask
If I enter my own API key, does the transcript stay private to my Mac?
No. The API key authenticates a request to the selected endpoint. The request still sends transcript text and formatting instructions there. Provider data-use terms depend on the specific account and agreement.
Does switching the provider off undo an earlier upload?
No. Turning off Cloud API changes future formatting requests. It does not delete content already processed by a provider; check that provider's retention and deletion terms.
Can a formatter fail without sending anything?
Some failures occur before a request, such as a missing key or invalid URL. Others happen after an HTTP request has reached the provider. A local fallback in the UI does not identify which occurred.
Related guides
- Does on-device mean no internet?
- Does SuprFlow use dictation for AI training?
- Choose local or cloud formatting
- What a no-training agreement should say
- What usage telemetry can reveal