The short version: prompts and responses are not stored. We keep token counts and timestamps because budgets have to be enforced somehow.
What passes through
ClaudeOpus proxies your requests to an upstream model provider and pipes the response back to your client. Prompts, system instructions, attachments, and generated content all travel through our servers in flight. None of it is retained beyond the life of the request.
What is never stored
The text of your messages, the attachments you send, the model's generated output, and any conversation history beyond a single round trip. If it came from you or the model, it is gone the moment the HTTP response closes.
What we do store
Usage metadata: which API key made the request, how many tokens came in and went out, the model requested, and the round-trip latency. This exists to enforce budgets, populate dashboards, and diagnose incidents. It does not include prompt or response content.
Account data
If you have a dashboard account, we keep the fields you typed in: email, name, hashed password. Resellers also store the hierarchy of sub-users they created. Nothing beyond what the dashboard itself shows you.
Upstream providers
Requests are forwarded to a third-party model provider to generate completions. That provider has its own privacy posture and log retention — review their policies if that matters for your workload.
Operational logs
Edge and server access logs (IP, timestamp, status code, path) rotate on a short window and exist only to investigate abuse and outages. We don't sell them, share them, or link them to prompt content.
Security
All traffic is HTTPS. API keys are hashed at rest where feasible and rotated easily. Anyone with your key can issue requests as you — revoke and reissue if it leaks.
Changes and contact
This page is the canonical version. Material changes will bump its date. For questions, reach the admin who issued your key.
— Questions? Your admin has the answers. —