⌗ privacy
Privacy.
last updated · 2026.05.04
What we collect
You set up an account with an email. The smoothtalk server stores creator profiles you configure (name, voice notes, banned words), per-fan metadata you record (display name, spend, kinks, free-text notes), the messages you paste in to get suggestions, and the suggestions you approve, edit, or reject.
What we do not collect
Your LLM provider key (Anthropic, OpenAI, xAI, OpenRouter) lives in the smoothtalk server config you control. We do not transmit it anywhere except the provider you chose. We do not scrape any chat platform on your behalf. We do not read messages you have not pasted in.
Why we keep approval data
Every approve, edit, and reject is the training signal that makes suggestions match your voice. It stays scoped to the creator profile it came from. You can wipe a creator at any time with a single click in the dashboard, which deletes their messages, fans, suggestions, and feedback.
Where it lives
Self-hosted: nowhere we control. Hosted plan: a Postgres database in a single region you select on signup. Backups encrypted at rest.
Sharing
We do not sell or rent your data. Any per-creator fine-tunes you run on the hosted plan are private to your account.
Pooled training set (opt-in, default-on)
When you opt in (default at signup, toggle on the bootstrap page and in app settings), the approved or edited replies you send are added to a pooled training set in PII-stripped form. We strip: fan handles, emails, URLs, phone numbers, dollar amounts, and any two-word capitalized names before insertion. Only your outbound reply text, the scenario class, and the voice template label are kept. No fan inbound messages ever leave your namespace. The pool powers a shared retrieval layer that surfaces high-performing replies across the network as additional in-prompt examples when relevant. You can toggle this off at any time and previously contributed entries are deleted on request.
Last updated
2026.05.10 — added pooled training set section.