Answers on sign-in, rollout, enforcement, cost, models and gateways for Claude Desktop, Claude Code, ChatGPT Desktop and Codex on Amazon Bedrock.
3P means the client sends inference to a third-party platform instead of the vendor’s hosted service. With Amazon Bedrock, Claude Desktop and Claude Code call models through your own AWS account, and people sign in through AWS IAM Identity Center.
Client settings choose the route; IAM and SCPs decide what is allowed.
ChatGPT Desktop and Codex use the built-in amazon-bedrock-runtime provider. Requests go from the local client straight to Bedrock Runtime with the person’s temporary AWS credentials, with no translation gateway in between.
The current generator puts provider requirements, new-thread model defaults and sign-in refresh in requirements.toml. AWS permissions decide which models can actually be invoked.
AWS-centric organizations that already govern identity, spend and audit in AWS and want AI desktop and coding clients inside that boundary.
Inference is billed by AWS on your account. Confirm licensing and commercial terms with the relevant vendors; AWS IAM, applicable SCPs and configured CloudTrail logging govern the Bedrock requests.
Bedrock requests use temporary AWS credentials from AWS IAM Identity Center. Identity Center can federate with Microsoft Entra ID, so people use their existing corporate credentials for AWS access. A client may also have its own initial app sign-in.
Claude Desktop has built-in Identity Center sign-in; managed setup for Claude Code, ChatGPT Desktop and Codex refreshes AWS SSO sign-in when credentials expire.
Generate the client files from one config.json: macOS managed profiles, a Windows .reg for Claude, an AWS CLI profile and Codex requirements.toml. On Windows, Codex uses a protected requirements.toml file, not registry policy. Distribute them through MDM or bootstrap delivery.
Finish by running the test plan to confirm sign-in, model calls and audit for each client and operating system in your rollout.
No. Claude Code availableModels limits the aliases shown in the model picker, not raw model IDs. A user can still request a model by its ID.
To limit models, scope the IAM permission set or attach an SCP, which AWS enforces regardless of client settings.
No. requirements.toml enforces the Bedrock provider, AWS profile, Region and disabled Runtime web search. Its new-thread section sets model and reasoning defaults that users can change, and there is no supported model allowlist in that setting.
Use IAM permissions or an SCP to decide which models are allowed.
The bedrock-governance SCP restricts direct model access in the target accounts or OUs, with temporary exceptions for Global OpenAI system inference profiles and the account’s project/default resource used by the Runtime Responses API. IAM must still scope access to approved models, and every request must satisfy applicable policies. The SCP does not grant access by itself.
For Claude, application inference profiles tagged per team support token metrics and cost allocation.
Per-profile tracking does not apply to ChatGPT and Codex in the same way; attribute their usage through the employee’s role session in configured CloudTrail or Bedrock invocation logs.
The bedrock-governance deployment configures a multi-region CloudTrail trail for Bedrock invocation events, with encrypted, versioned storage and a 30-day S3 audit-log lifecycle. Confirm retention and logging scope for your deployment.
CloudWatch anomaly alarms watch input and output token usage on configured inference profiles.
The managed setup uses approved Global or US Geo Runtime inference profile IDs rather than application inference profiles. Confirm current availability by model, AWS account and Region before choosing the managed default.
Which of these a person can call is set in IAM.
Yes. coding-agent-3p generates the client configuration and bedrock-governance deploys the account guardrails; both are public on GitHub and free to use.
ASCENDING offers a separate paid implementation service with a 3-day rollout plan, subject to agreed scope and prerequisites.
No. Claude, ChatGPT and Codex clients call Bedrock directly with their own AWS credentials, and IAM, applicable SCPs and configured CloudTrail logging apply to those calls.
A gateway is optional for additional policy, filtering or attribution on traffic routed through it.
Jarvis is an optional governance layer for model and tool traffic routed through it, adding identity-based model policies, guardrails and team attribution. Jarvis MCP Gateway governs tool access for agents.
Neither is required to run Claude, ChatGPT or Codex on Bedrock.
3P on Amazon Bedrock is delivered by ASCENDING, an AWS Partner with the AI Services Competency. These rows show where each governance question is answered: on the device or in your AWS account.
| Governance question | Where it is answered |
|---|---|
| Which provider does each client use? | Managed configuration on the device: Claude managed policy, or requirements.toml for ChatGPT and Codex. |
| Which models are allowed? | IAM permission sets or an SCP. Client allowlists are not a boundary. |
| How do people sign in? | AWS IAM Identity Center, optionally federated with Microsoft Entra ID. |
| Are direct model calls blocked? | The SCP restricts direct access, with temporary exceptions for Global OpenAI system inference profiles and project/default. IAM must still limit access to approved models. |
| Who used what? | Configured CloudTrail and Bedrock invocation logs, by role session. |
| How is cost attributed? | Tagged application inference profiles for Claude; role-session audit for ChatGPT and Codex. |
The same 4 steps apply to every client. Each step can be verified before the next, so a pilot group can be live before a wider rollout.
Each client family has its own page with setup, governance and rollout detail beyond the answers above.
First-party documentation and the open-source repositories behind the answers on this page.
Bring your rollout plan and we'll walk through the controls with your team.
