LLM routing, caching and guardrails
How the gateway picks a provider connection for each model call, and what it does not do yet.
If you don't see this feature in your console, ask your SecureAuth contact to
enable the llm_gateway feature flag for your organization.
With more than one provider connection, each model call has to reach one that can serve it. The AI gateway routes each call to a connection that serves the requested model and that your model rules allow.
Routing
The gateway considers the connections that use the call's API format, OpenAI or Anthropic, and the same kind of credential: a shared API key or a user sign-in. It keeps the ones whose Allowed models include the requested model, and refuses the call if none does. Your model access rules then narrow those connections, because a rule can match model providers. A request that names more than one model is checked against your rules for each model.
Caching and guardrails
The gateway does not cache model responses. It does not filter or redact the content of model calls. Data Protection filters apply to MCP tool responses, not to model calls.
To control which models each agent and user reaches, see Model access rules.