AWS shows how to route Codex through a company-controlled model gateway

Companies can route OpenAI Codex requests through a LiteLLM gateway running in their own AWS account, according to a new deployment guide from AWS. The architecture keeps Codex’s coding loop on the developer’s workstation while sending model inference through Amazon Bedrock under centralized controls.

LiteLLM runs on Amazon ECS with AWS Fargate and exposes the Responses API expected by Codex. It becomes the shared point for authentication, model routing, budgets, rate limits, and telemetry. The guide also validates streaming, function calling, and semantic continuation so multi-step coding tasks can continue across model requests instead of losing state at the gateway.

This separation gives administrators visibility into model consumption without moving Codex’s local file reading and approved tool execution into the cloud. Scoped identities can attribute usage to particular users or teams, while budget and rate policies limit how much traffic they generate. AWS has published the implementation in a guidance repository.

The self-operated path adds infrastructure that must be deployed, secured, monitored, and upgraded. AWS therefore also compares direct IAM Identity Center access and a managed Portkey gateway, which may fit teams that do not need LiteLLM’s flexibility. None of these options replaces Codex’s local sandbox and approval settings; the gateway governs model access, not every action the coding agent can take on a workstation.