Ask HN: Would you pay for an AI API that guarantees zero prompt retention?

4 points by mhrnik 23 days ago | 7 comments

i am just tired of looking for a reliable open-source model AI API with ZDR and don't even save my prompts. I'm thinking of a service for people like me who want to use it.

idea is a simple OpenAI-compatible API:

Access to open-weight models

zero retention of prompts and completions

no training on data

only retain metadata required for billing and operations: request ID, model, input/output token counts, latency, timestamp, etc.

no request/response in logs

Just here to check interest. I am not 100% sure if I can get compute to do it or not (may be i can just try with my 5090 first with a small model), but I just want to check and talk with people rather than just thinking in my head.

ljlolel 23 days ago |

My router provides that! TrustedRouter.com we have many end to end encrypted models, and providers including Tinfoil but also others

Importantly i have confidential compute so you can verify the end to end process of not logging the data -- at all.

https://trustedrouter.com/blog/how-confidential-computing-pr...

mhrnik 21 days ago | |

thanks, checking

selfhoster11 23 days ago |

Tinfoil etc offer TEE inference that works well, directly or via API routers.

mhrnik 23 days ago | |

thanks. tinfoil [dot] sh looks promising. trying it.

cregy 23 days ago |

Openrouter offers that already on a few models

minimaxir 23 days ago | |

Specifically, you can set an account setting "only do ZDR" and it will error if you try to use an endpoint that doesn't do it.

mhrnik 23 days ago | |

Tried and not reliable.

last day only I visited to try qwen3.8-max, and there is only one provider alibaba and personally I don't prefer them.