Every prompt is stripped before it reaches a model outside your walls. No foreign vendor holds your client’s name.
Client names, IPs, credentials, tokens, emails, card numbers: stripped before a single byte leaves the premises.
Anthropic today, OpenAI tomorrow, local only forever if you want it. No model vendor owns your workflow.
If a prompt still looks sensitive after sanitization, or you tell it to, it never leaves. A local responder answers instead.
Every query logged: provider, model, whether it was sanitized, a hash of prompt and response. Never the raw content.