Privacy by Design
Local transcription is not a privacy feature — it's the architecture. Audio is processed on your device and never transmitted to any server.
Core privacy principles
On-device processing
Transcription, speaker labels, translation, search, and chat all run on your computer's CPU or GPU. Audio is read, processed, and stored locally, and your audio and transcript content never travel over the network. First-run model downloads, licence checks, and update checks are separate from processing and carry none of it.
Minimal data collection
Kajo's servers hold your licence key, the email address it was sent to and Creem's order ID, the install IDs and session records of the computers that use it, feedback you choose to send (kept a year), waitlist sign-ups until launch, hashed network and browser identifiers used for abuse limits (kept at most a month), and request logs (kept 7 days). Product analytics (PostHog, EU) run only if you tick “Help improve Kajo”, which is off by default, and go straight to PostHog, never through Kajo's servers. Crash reports (Sentry) are sent only if you agree when the app asks at first run, where the choice starts unticked; change it any time with “Send crash reports” in Settings → Privacy.
No cloud AI path
Summaries, translation, and chat all run on a local model on your device. There is no cloud AI feature to enable and no cloud subprocessor that receives your content — Kajo's servers never see audio, transcripts, or library content.
Verified downloads and updates
Every model the app downloads is checked against a pinned SHA-256 checksum before it is used, and on a Mac or Linux PC every update is checked against a signed (Ed25519) release manifest before it installs. A file that does not match is refused. On Windows, updates come through the Microsoft Store, which signs every package it delivers.
Protected in transit, private on disk
Everything the app exchanges with Kajo's servers travels over TLS, and licence certificates are signed with Ed25519 and verified on your machine. Transcripts, chat history, and search embeddings live in a local SQLite database (`kajo.db`) with no database key: your OS account and disk encryption are what protect it, so turn on FileVault, BitLocker, or full-disk encryption. Audio — your originals and Kajo's compact playback copies — stays on your disk under the same permissions. None of it is ever uploaded.
Where your data lives
| Data type | Where it lives | Protection |
|---|---|---|
| Audio files (your originals and Kajo's compact playback copies) | Your device only | OS file permissions — use disk encryption. Never transmitted. |
| Transcripts, library, chat history, embeddings | Your device only (`kajo.db`) | Plaintext SQLite under your OS account — use disk encryption. Never transmitted. |
| License key and email address | Cloudflare D1 (EU jurisdiction) and your OS keychain; a purchase email is also held by Creem, and Resend keeps licence emails for 30 days | Encrypted at rest by Cloudflare (AES-256); your local copy is protected by the OS keychain. Licence certificates are signed with Ed25519. |
| Payment details | Creem only | Creem processes payment details; Kajo never receives raw card data. |
Sub-processor list
Every third-party service that processes data for Kajo — none of them receives audio, transcripts, or library content. With their DPAs: /legal/subprocessors.
Creem
Privacy policy ↗Sells the Lifetime licence as our contractual reseller and merchant of record; processes checkout, payment, applicable taxes and receipts. Creem’s privacy notice describes its role as controller, joint controller or processor depending on the processing activity
Cloudflare
Privacy policy ↗Hosting and delivery: the kajovoice.com website, the installers, the update feed and model downloads (Workers and R2), and the licensing API with its database of licences, sessions, feedback and waitlist sign-ups (Workers and D1); authoritative DNS; inbound email routing for kajovoice.com; and alert emails to Kajo's owner
Hugging Face
Privacy policy ↗Fallback source of on-device model weights: every model downloads from Kajo's asset CDN, and from Hugging Face when the CDN is unavailable; Hugging Face handles these requests as an independent controller
PostHog
Privacy policy ↗Pseudonymous usage events from the desktop app — which features are used, how long they take, and error details — sent only if you tick “Help improve Kajo” (off by default); never audio, transcripts, file names, or chat text
Resend
Privacy policy ↗Transactional email: the license-key email sent after purchase, and the waitlist confirmation and launch emails
Sentry
Privacy policy ↗Crash reports from the desktop app, sent only if you agree when the app asks at first run (the choice starts unticked); change it any time in Settings → Privacy
Privacy posture
Kajo is built so that nothing needs uploading: your audio and transcripts are processed on your device and never sent to our servers.
- On-device processing: Whisper large-v3 turbo, the speaker model, and the local language model all run on your machine — audio never leaves it.
- No-upload architecture: the app has no code path that sends audio, transcripts, or library content to a server, so there is nothing to switch off.
- No DPA for content processing: your recordings never reach Kajo, so Kajo does not process them on your behalf (/legal/dpa).
Security disclosure
Found a vulnerability? Please disclose responsibly to:
security@kajovoice.comLifetime customers can request a security questionnaire response at security@kajovoice.com.