Does Soniteq Train on My Data?
Direct answer: No. Your audio never leaves your machine, so it cannot be used to train any model — ours or a music-generation service like Suno or ElevenLabs. Kai works on workflow metadata: project names, deadlines, delivery status. That text goes to our own proxy and, under the AI providers’ standard API terms, is not used to train their models either.
The part that is absolute
No audio ever leaves your machine. Not stems, not mixes, not masters, not reference files — at any tier, on any plan, whether or not cloud sync is on. There is no code path that uploads audio content, and no integration with any music-generation service. A repository-wide search for the obvious ones — Suno, Udio, MusicGen, ElevenLabs, Stable Audio, Riffusion, AudioCraft — returns nothing, because no such connection exists.
This is the claim worth checking first, because it is the one that cannot be softened by a change of terms somewhere else. Your catalogue cannot train a music model that never receives it.
What does leave, and under what terms
Kai’s cloud requests carry text, and only text:
- Names of projects, albums, contacts, releases and opportunities
- Dates, statuses, priorities and release-readiness scores
- Your composer profile and learned memory facts
- The last 8 chat turns, truncated to 220 characters each
- Note snippets — up to 600 characters from as many as 15 notes, and the full text of a note when you explicitly ask Kai to summarise or rewrite it
- Your export folder path, sent with the delivery digest
Those requests go through Soniteq’s own proxy rather than from your machine to a model vendor directly. Soniteq does not train on any of it, does not sell it, and does not build behavioural profiles from it.
On the providers’ side: the standard API terms used by commercial AI vendors exclude customer API traffic from model training by default — this is the distinction between their API products and their consumer chat apps. Providers may still retain API data for a limited period for abuse monitoring unless a zero-retention agreement is in place, so this page claims no training, which is the commitment that holds, rather than no retention, which would depend on an arrangement we are not asserting here.
Local is the default
Kai’s engine mode defaults to local, and the Local tier never contacts the network at all. Cloud models handle heavier requests when they are available, and if a cloud call fails the chain falls back to the local engine rather than erroring. Cloud AI is a processing step for a request you initiated — not a collection channel that runs in the background.
Your Kai profile, memory and beliefs stay on your device. Secrets stay in the macOS Keychain.
Why this matters
Creative work is leverage. Handing a catalogue of unreleased music to a model vendor is an irreversible decision, and most tools bury that permission in their terms of service. Soniteq states its position plainly so it can be checked rather than assumed.
Where this is documented
The full commitment, including retention and deletion behavior, is on the Data Philosophy page.
Related answers
- Is Kora local-first and offline?
- Is my data secure in Kora?
- What is creator-first AI?
- What is Soniteq?