Privacy · storage · deletion
Local AI, explained precisely.
MumbleFlow is built to keep customer speech and writing on the Mac. This page explains what “local-first” means, where network access still occurs, how stored data is protected, and when it disappears.
Apple Silicon · macOS 26+ · English · Models download during setup
On-device processing
Your content is not an inference-service input
FluidAudio voice detection, Parakeet speech recognition, Apple Foundation Models or optional Qwen polishing, meeting diarization, summaries, embeddings, hybrid retrieval, and grounded answers execute locally. The Mac app has no MumbleFlow account, ambient buffer, automatic meeting recording, analytics SDK, cloud content server, or cloud sync.
Local processing does not control the app that receives your final text. If you dictate into a cloud document, shared message, browser form, or synced notes app, that destination's policies apply after insertion. Follow the complete voice-to-text data flow.
Lifecycle
Keep what is useful, expire what is not
Dictation audio
Normally exists only in memory and is discarded immediately after processing. Opt-in diagnostics are a separate, visible exception.
Dictation text
Encrypted history expires after seven days by default. Disable history, pin specific entries, or clear unpinned entries at any time.
Meeting audio
Separate microphone and computer-audio channels are encrypted during recording and expire after 30 days unless deleted sooner.
Meeting knowledge
Transcripts, notes, citations, and confirmed speaker profiles remain locally until you delete the associated data.
Diagnostic sample
Off by default. When enabled, only the newest test WAV and raw/final comparison are retained, and each new test replaces the previous one.
Model assets
Speech assets and Qwen contain no customer content. The optional Qwen fallback can be removed directly from Settings.
At-rest protection
Different data gets separate protection
- 01
Customer records
Dictations, transcript content, meeting metadata, sources, names, and speaker embeddings are sealed field by field with AES-GCM and context-specific associated data.
- 02
Meeting recordings
Audio is appended to sequence-authenticated AES-GCM journals as the meeting happens; MumbleFlow does not leave a plaintext hour-long recording file behind.
- 03
Search indexes
FTS5 receives opaque HMAC tokens instead of recoverable words, and the serialized USearch vector archive is encrypted separately.
- 04
Secrets
Content keys and any future connector credentials use This-Device-Only Keychain items.
These are specific implementation properties, not a blanket certification. The current SQLite container is not advertised as SQLCipher-encrypted; the sensitive content inside it is individually protected.
Internet access
Local-first does not mean zero network traffic
- Buying and downloading the app uses the website and Stripe checkout.
- Required versioned speech assets download during setup and are verified before use.
- The optional Qwen3 1.7B fallback is a user-initiated download of roughly 930 MB; eligible personal vocabulary can trigger a separate roughly 98 MB acoustic rescoring model.
- Update downloads require network access when an update channel is enabled. Automatic update checks are not enabled in the current package.
- External Gmail, Calendar, and Slack traffic would occur only after those connectors ship and a user explicitly connects them. They are not enabled in the current download.
The previous marketing website used Vercel Analytics. This redesigned website does not include that integration. Website hosting and Stripe checkout remain separate from the MumbleFlow Mac app, which contains no analytics SDK.
User controls
Delete without filing a support ticket
Dictation History can be disabled or cleared for unpinned entries. Personal vocabulary supports edit, remove, and clear-all controls. Each meeting lets you delete retained audio immediately or erase the complete meeting, transcript, notes, and search sources. Optional Qwen assets can also be removed in Settings.
For help with a complete reset, visit MumbleFlow Support or email contact@aura-technologies.co.
Questions
Frequently asked questions
Does MumbleFlow upload my dictation or meeting audio for AI processing?
No. Speech recognition, writing cleanup, speaker analysis, meeting summaries, embeddings, meeting search, and generated answers run on your Mac. Model files and app updates may be downloaded, but customer audio and text are not sent to a hosted MumbleFlow inference service.
Does the MumbleFlow Mac app have an account or analytics?
The Mac app has no product account, analytics SDK, cloud content server, or cloud sync. The website and Stripe checkout are separate services from the desktop app. This redesigned website does not include the previous website’s Vercel Analytics integration.
Is the MumbleFlow database SQLCipher encrypted?
The current app does not claim whole-container SQLCipher encryption. Sensitive records are encrypted field by field with AES-GCM, search tokens are opaque HMAC values, meeting audio uses authenticated encrypted journals, and the serialized vector index is encrypted separately.
Can I delete MumbleFlow data immediately?
Yes. You can clear unpinned dictation history, remove personal dictionary entries, delete a meeting’s audio, erase a meeting and its transcript, and remove optional Qwen model files from Settings.
Keep exploring
The rest of MumbleFlow
All features
A complete tour of dictation, Oat meeting notes, search, personalization, privacy, and controls.
AI dictation for Mac
Natural voice typing that removes fillers, catches corrections, formats lists, and inserts finished text.
Oat AI meeting notes
Reviewable, actionable meeting notes with live catch-up, speaker memory, timelines, and source citations.
Meeting transcription
Separate microphone and computer audio, readable speaker turns, timestamp playback, and grounded notes.
Local AI search
Private hybrid search and grounded answers that take you directly to the supporting source.
About MumbleFlow
Why MumbleFlow exists, how it is built, and the product principles behind it.
Looking for technical detail? Read the local speech-to-text guide.