@9hills: At first, I didn't really get what Anthropic was trying to do — why detect Claude Code using third-party providers? What's the big deal? Later I realized how clever Anthropic is: by modifying the date format in the system prompt and comparing it with their own logs, they can precisely identify proxy station IPs and accounts. As long as your local configuration points to a third-party URL but the request ultimately goes back to Anthropic, it must be a proxy station. So they can batch-ban those proxy stations.

I actually didn’t initially understand what Anthropic’s goal was—detecting third‑party providers for Claude Code, what’s the big deal?

Later I realized how clever this is. By modifying the date format in the system prompt and comparing it with Anthropic’s own logs, they can precisely identify relay station IPs and accounts. If you configure a third‑party URL locally but the requests still go back to Anthropic, it must be a relay station. This allows them to ban relay stations in bulk.

Similarly, many companies—whether for distillation or for using the Claude model—can’t possibly give each employee their own account, so they typically use self‑built relay station solutions. Even if the backend is AWS, it can still be precisely identified.

Finally, because the detection is injected into the prompt and involves no network upload activity, it’s hard to spot. Even if your enterprise’s Claude Code runs in an internal network environment, as long as the model requests eventually go back to Anthropic, they will be discovered.