Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableAWS Machine Learning Blog→ original

ChatGPT and Codex Now Run in AWS GovCloud: How Government Agencies Can Lock Down the Network

US government agencies cannot send code to OpenAI's regular servers, so AWS has published a workaround: the ChatGPT app and the Codex coding agent work with the models through Amazon Bedrock in AWS GovCloud (US-West). Two configuration files and an allowlist with a single network address are enough. You pay for tokens, not per-seat licenses. Security teams get a ready-made template they can test before a wide rollout.

AI-processed from AWS Machine Learning Blog; edited by Hamidun News
ChatGPT and Codex Now Run in AWS GovCloud: How Government Agencies Can Lock Down the Network
Source: AWS Machine Learning Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceAWS Machine Learning Blog ↗
Editor-in-Chief
Key metric #1Only the Bedrock domain is allowed, while api.openai.com and chatgpt.com are blocked
Zero data retention moderequests and responses are not written to disk
The guide alone is not enoughthe organization must authorize the workload itself

Large companies and government agencies often cannot simply roll out an AI assistant for developers. Code, documents, and internal data cannot be sent to external servers. Requests need to flow through clearly defined endpoints, stay in the right region, and be controlled at the network level. AWS has published a guide that covers this scenario for the AWS GovCloud (US) cloud: the ChatGPT desktop app with the Codex agent can now be connected to OpenAI models through Amazon Bedrock.

What was actually launched

The app has two modes. The first, Work, handles large tasks: planning, research, file work, and finished deliverables. The second, Codex, is built for technical work: it edits code, runs tests, and reviews pull requests. Codex is available in the app, in the command line, and as plugins for Visual Studio Code, JetBrains, and Xcode.

Model requests go through the Responses API on Amazon Bedrock. This uses the Bedrock Mantle endpoint in the us-gov-west-1 region. If you have data residency requirements, processing stays in the AWS region you choose. You pay for tokens, with no per-seat licenses and no commitments per developer. Zero data retention (ZDR) mode is also supported: AWS does not write requests and responses to persistent storage and does not pass them to the model provider.

How it is configured

The setup is straightforward. The app runs on a computer managed by the organization and uses local AWS credentials. Access is determined by two things: the policies on the machine itself and the permissions of the AWS identity.

Authentication is configured first. You can use an Amazon Bedrock API key or the standard AWS SDK credential chain, including SSO through IAM Identity Center. If the AWS_BEARER_TOKEN_BEDROCK variable is set, the token takes priority over the profile. The example uses the codex-il5 profile. Secrets must not be stored in TOML files, and the profile files and SSO cache must be protected from writes by the agent itself.

Next, two files are created. requirements.toml sets administrative constraints that the user cannot override. It allows sign-in only through the API, restricts sandbox modes to read-only and workspace-write, and disables the browser, computer control, remote control, plugins, memory, and image generation. The network allowlist passes only bedrock-mantle.us-gov-west-1.api.aws, while OpenAI and ChatGPT addresses, including telemetry and updates, are explicitly blocked. config.toml points to the amazon-bedrock provider, an approved model, the region, and the profile, and turns off analytics, feedback, and telemetry. The files must be in place before the first launch, or the app will ask for an OpenAI account sign-in.

After launch, confirm that amazon-bedrock is listed as the provider in the app settings or in the /status tab of the command line.

Limitations and what to do

AWS states plainly that GovCloud and this guide do not, by themselves, authorize a workload. The organization must check the model, services, region, endpoint, workstation protection, repositories, and data flows against its own security plan. Because the connection goes directly through AWS, ChatGPT workspace rules and user roles do not apply here. All protection rests on machine policies and IAM permissions.

The guide names four common problems. The app asks for an OpenAI account: check the file paths and names, quit all ChatGPT and Codex processes, and relaunch. The endpoint is reported as unsupported: confirm the app version and model with OpenAI, since replacing the URL alone is not enough. The wrong identity is used: check whether the profile is overriding the API key. The restrictions did not apply: after restarting, review the full effective requirements, since the API sign-in setting alone does not enable the capability policy.

If validation fails, the rollout is halted, devices are returned to the previous policy, and temporary keys are revoked.

The practical takeaway for security teams: start with a small pilot, document the validation results, and only then expand access. The ready-made domain allowlist can save weeks of approvals, but the decision to use it still rests with your security team.

Story

Key Questions Answered

What was actually launched?

The app has two modes. The first, Work, handles large tasks: planning, research, file work, and finished deliverables. The second, Codex, is built for technical work: it edits code, runs tests, and reviews pull requests. Codex is available in the app, in the command line, and as plugins for Visual Studio Code, JetBrains, and Xcode. Model requests go through the Responses API on Amazon Bedrock. This uses the Bedrock Mantle endpoint in the us-gov-west-1 region.

How it is configured?

The setup is straightforward. The app runs on a computer managed by the organization and uses local AWS credentials. Access is determined by two things: the policies on the machine itself and the permissions of the AWS identity. Authentication is configured first. You can use an Amazon Bedrock API key or the standard AWS SDK credential chain, including SSO through IAM Identity Center. If the AWS_BEARER_TOKEN_BEDROCK variable is set, the token takes priority over the profile. The example uses the codex-il5 profile.

What are the key details on Limitations and what to do?

AWS states plainly that GovCloud and this guide do not, by themselves, authorize a workload. The organization must check the model, services, region, endpoint, workstation protection, repositories, and data flows against its own security plan. Because the connection goes directly through AWS, ChatGPT workspace rules and user roles do not apply here. All protection rests on machine policies and IAM permissions. The guide names four common problems.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…