One key per protocol
Some relays hand out one key for Anthropic and another for OpenAI. Each key in a provider's Accounts list can now say which protocol it works with: Any protocol, Anthropic only, Chat Completions only or Responses only. The picker shows when the provider has more than one endpoint.
The gateway uses each key on its own endpoint only, and puts first the key that fits the request:
- Claude models go to the Anthropic key, and GPT/Codex models go to the OpenAI one. The gateway translates between protocols when it has to, so it doesn't matter which agent sent the request.
- For other models, the gateway prefers the key that speaks the agent's own protocol, so nothing gets translated that doesn't need to be.
- Token counting for Claude Code uses the Anthropic key.