ChatGPT desktop app can now run on Bedrock inside AWS GovCloud, with OpenAI sites blocked

Amazon Web Services says the ChatGPT desktop app, including its Codex coding agent, can now send model requests to Amazon Bedrock in AWS GovCloud (US). A walkthrough published October 30 targets agencies that need inference to stay under GovCloud identities and endpoints.

A locked-down setup that blocks OpenAI’s own servers

The app has two modes: Work, for multi-step tasks that end in a finished result, and Codex, for writing, testing and reviewing code. According to AWS, Codex also runs through the CLI and IDE integrations for Visual Studio Code, JetBrains and Xcode. Requests go through the Responses API to the Amazon Bedrock Mantle endpoint in us-gov-west-1, at bedrock-mantle.us-gov-west-1.api.aws.

The guide relies on two files that must exist before the first launch. An administrator-controlled requirements.toml sets limits users cannot override. A user-level config.toml selects amazon-bedrock as the model provider and names an approved model ID and AWS profile. The example profile is codex-il5.

The baseline policy is strict. It allows only API-mode sign-in and the read-only and workspace-write sandbox modes. It turns off browser use, computer use, remote control, plugins, memories, image generation, in-app updates and MCP servers. Network access is limited to the Bedrock endpoint, while a list of OpenAI and ChatGPT domains, including api.openai.com and chatgpt.com, is explicitly denied. Analytics, feedback and telemetry are disabled.

Authentication works two ways: an Amazon Bedrock API key supplied as AWS_BEARER_TOKEN_BEDROCK, or the standard AWS SDK credential chain, which includes AWS IAM Identity Center SSO. The bearer token wins when both are present. AWS says secrets should stay out of the TOML files.

Per-token billing and what the source leaves open

AWS says customers pay per token, with no seat licenses or per-developer commitments, and that zero data retention is supported as a configurable mode. For data residency, processing stays within the Bedrock Region the customer selects. The post gives no prices, no model names and no list of supported app versions.

It also carries a caution. Using GovCloud and following the guide “do not, by themselves, authorize your workload”, and each organization must check the model, Region, endpoint and workstation controls against its own security plan. Another warning says changing the URL alone does not establish support, so customers should confirm the exact desktop release, endpoint and model with OpenAI. The post is a configuration guide, not a certification, and the security claims are AWS’s own.

One practical consequence: because the app talks to AWS directly, AWS notes that ChatGPT workspace rules and user roles do not govern it. Control shifts to device policy and the IAM permissions of the identity in use.

Part of a wider GovCloud push for OpenAI models

The guide landed alongside other GovCloud news the same day. Amazon separately said Bedrock now supports OpenAI’s GPT-5.6 models, Terra and Luna, on the bedrock-runtime endpoint in GovCloud (US), through the Responses, Converse and Chat Completions APIs. It also added Cross-Region Inference, which spreads requests across both GovCloud Regions while keeping data in the US. That notice does not name the Regions or give pricing, and it does not say which model the desktop guide expects.

A second notice covers Daybreak Red, GPT-5.6 Cyber, available to eligible customers in GovCloud (US-East) and US East (Ohio). Access requires enrollment in OpenAI’s Daybreak program and approval through an AWS account team. The desktop guide, by contrast, is pinned to US-West.

Taken together, the posts suggest AWS is assembling a full stack for agencies: models, regional failover and now a client. What to watch is whether the desktop guide is later extended to the Cross-Region setup, and whether agencies publish approvals for the configuration, since the guide itself leaves authorization to each organization. Related AWS agent tooling is covered in our report on AgentCore live previews of coding agents.