AWS routes ChatGPT desktop through Bedrock in GovCloud

AWS has published a supported configuration for routing inference from the ChatGPT desktop app and Codex through Amazon Bedrock Mantle in AWS GovCloud (US-West). Government-cloud organizations can use AWS identities, endpoints and regional controls for local coding workflows, but the configuration does not itself grant an authorization to operate.

In this setup, Codex runs on the user’s device and sends model requests to Amazon Bedrock’s OpenAI-compatible Responses API. The inference path leads to the configured Bedrock endpoint, not OpenAI’s hosted Responses API. Device policy, the effective AWS identity, IAM permissions and approved network routes are therefore the main control points for this API-mode deployment.

A rollout begins with an approved GovCloud account, model, identity, credentials, Region and endpoint. Before the app launches, administrators deploy a protected requirements.toml. Users cannot override this policy layer, which can restrict login to API mode, limit sandbox and approval settings, disable features including browser control, computer use, apps, plugins, MCP servers, image generation and memories, and constrain permitted network destinations.

Within those limits, a separate config.toml selects the Amazon Bedrock provider, model, endpoint, Region and named AWS profile. The published example pairs the Bedrock Mantle hostname for us-gov-west-1 with the same AWS Region. Its managed network policy allows that hostname while denying listed OpenAI and ChatGPT destinations. Administrators must keep credentials outside both the configuration file and folders Codex can modify.

Authentication is only the first gate. Amazon Bedrock’s endpoint documentation says the associated IAM principal also needs the required inference permission: bedrock-mantle:CreateInference for the Mantle endpoint. The client can use a Bedrock API key or AWS credentials, including named profiles and IAM Identity Center sign-in, but those credentials do not expand what IAM authorizes.

Regional behavior follows the selected endpoint. Bedrock Mantle endpoints are Region-specific and do not support cross-Region inference, a regional-control issue also relevant to AWS’s Canadian Bedrock access to OpenAI GPT models. AWS’s Responses API documentation says stored Mantle response data remains in the Region that received the request.

Retention requires an explicit choice. AWS describes zero-data-retention support for the configurable route, but the Responses API documentation says storage is enabled by default. When store is true, Amazon Bedrock retains the input and output for 30 days. Organizations seeking no Bedrock retention must set store to false on every request or set the account retention mode to none.

Technical availability remains separate from an agency’s authorization decision. The GovCloud configuration guide directs organizations to review the model, AWS services, Region, endpoint controls, devices, repositories, credentials and data flows against their system security plan with security and authorizing officials. AWS’s GovCloud compliance documentation describes the environment and compliance programs, but does not make a particular workload authorized.

Before rollout, the guidance calls for checking the active policy, provider, model, AWS identity and endpoint; running an approved sample workflow; confirming that users cannot override restrictions; and testing invalid credentials and route failures. If validation fails, AWS directs administrators to stop the rollout, restore the previously approved provider and policy settings, revoke temporary credentials or API keys, verify that Bedrock connectivity is gone and document the failure before trying again.