AWS explains local ChatGPT deployment in GovCloud

AWS’s deployment guide describes a local ChatGPT desktop setup that sends model requests to Bedrock Mantle in the US-West GovCloud region. It targets organizations that need AWS identities and regional controls for inference, meaning the model’s processing of requests. Administrators must put configuration and managed requirements in place before launch. Workstation policies and AWS permissions govern this route; ChatGPT workspace roles do not. The organization still has to review and authorize its workload.

OpenAI’s client documentation explains authentication separately. The local client first checks for a Bedrock API key, then uses the AWS SDK credential chain if no key is available. A selected profile therefore does not necessarily determine the active identity. The model must be available in the chosen region and accessible to that identity. Verification includes checking the provider in the client: a successful short reply alone does not prove the route. Hosted Work on the web and Codex cloud are outside this configuration.

The underlying Responses API documentation also distinguishes two Bedrock endpoints. AWS recommends Runtime for new applications when the required model and features are supported. Mantle remains relevant for capabilities such as background inference and server-side tools. These endpoints are not interchangeable simply because both accept Responses requests. Model and API support must be checked for the specific endpoint before choosing a route.