Own hardware

The AI model of elmtree chat runs on elmtree's own computers in our premises in Berlin-Kreuzberg. Nobody except us has access to them. There is no external AI provider, not even as a fallback: if our computers are busy or unreachable, you get a notice and can send the message again.

How your request travels

Your message goes to our server at Hetzner in Germany. From there we forward it over our own encrypted connection (WireGuard) to our computers. No service provider that processes the content sits in between.

What our computers receive

Our computers receive the content of your request and the conversation context needed for the answer. They do not receive an account ID, name, email address, IP address, or session identifier.

If you use pseudonymization, the model sees the pseudonymized version with tokens instead of the detected original values.

What our computers do not do

• They do not store your requests and answers and do not write them to logs. Only technical data such as time, duration, and token count is logged, without any link to your account.
• To answer follow-up messages faster, they keep parts of the conversation in memory for a short time – separated per account and never on disk.
• We do not analyze your chats and do not train models on them.

What you should still keep in mind

Even on our own hardware, data minimization beats any promise. Do not include personal data that is not needed for the task. For documents with sensitive information, use the preview and pseudonymization.

Web search is the only path on which parts of your request go to an external service: a search provider receives the filtered search query (Data minimization). Turn web search off when the task does not need it.