3P means the client sends inference to a third-party platform: here, Amazon Bedrock in your own AWS account, with people signing in through AWS IAM Identity Center. Client settings choose the route. IAM policies and service control policies (SCPs) decide what is allowed.
A first-party (1P) deployment uses the vendor’s hosted accounts. A third-party (3P) deployment keeps familiar desktop and command-line clients but sends inference through your AWS account. 3P suits organizations that already govern identity, spend and audit in AWS.
| Control | First-party plans | 3P on Amazon Bedrock |
|---|---|---|
| Billing | Vendor plan or enterprise agreement | Inference consumption billed by AWS; confirm separate licensing and implementation terms. |
| Sign-in | Vendor account, with SSO configured in the vendor organization | IAM Identity Center issues temporary AWS credentials and can federate with Microsoft Entra ID. |
| Data location | Vendor-hosted service | Inference runs through Amazon Bedrock, subject to your chosen routing scope and AWS policies. |
| Admin controls | Vendor admin console | Managed client policy, IAM, SCPs and CloudTrail. |
Desktop chat apps and coding agents run on employee devices, so governance works in two places. On the device, configuration chooses the route. In your AWS account, policy decides what is allowed.
Managed settings point each client at Bedrock, set the AWS profile and approved defaults, and renew sign-in when credentials expire. They choose the route but cannot enforce limits.
IAM and SCPs authorize Bedrock requests. Configured CloudTrail events and application inference profile alarms provide visibility; client settings cannot override AWS authorization.
Calls reach the controls in your AWS account. IAM and the SCP define access boundaries; the trail, alarms and dashboard make usage visible.
A person signs in and the client calls an inference profile, which routes to the model. For Claude, the SCP requires application inference profiles and denies direct foundation-model calls.
Claude usage and cost attach to a named, tagged resource. ChatGPT and Codex use Runtime system profiles under the module’s temporary OpenAI exception, with usage attributed by role session.
The IAM permission set authorizes which models and profiles a person may invoke. The SCP limits direct foundation-model invocations across target accounts or OUs.
The module currently exempts Global OpenAI system profiles and the default project. This exception does not approve every model: scope IAM permissions to the models your organization allows. Neither control is a spending cap.
Only from role sessions authorized by IAM.
Restricted by the SCP, subject to its documented OpenAI exception.
Client settings cannot grant access.
Terraform creates an application inference profile for each approved Claude model. Usage and cost attach to that named resource, and per-team tags support cost allocation.
Profiles do not grant access on their own; IAM authorizes calls. ChatGPT and Codex use Runtime system profiles instead.
Example application inference profile
Example application inference profile
Example application inference profile
A multi-region CloudTrail trail captures configured Bedrock invocation events in encrypted, versioned S3. The module applies a 30-day S3 audit-log lifecycle. CloudWatch anomaly alarms watch input and output token usage per application inference profile.
The trail records calls rather than blocking them. Alarms are not a chargeback dashboard or a hard cap.
cloudtrail · InvokeModel
cloudwatch · InputTokenCount anomaly
Use CloudWatch token metrics from application inference profiles to build reporting by model and team. Team tags support cost allocation; the module supplies metrics and alarms, not a ready-made dashboard.
Usage is shown per inference profile. ChatGPT Desktop and Codex calls are attributed by role session instead.
Two repositories cover client configuration and AWS account guardrails. The code is free; ASCENDING delivers implementation, configuration and support as a paid engagement.
Fill in one config.json and it renders the managed configuration for every client and operating system, with an admin guide and a test plan to verify the rollout.
A macOS .mobileconfig and a Windows .reg with built-in IAM Identity Center sign-in.
awsAuthRefresh for sign-in and modelOverrides mapping aliases to Bedrock IDs.
requirements.toml pins the Bedrock provider and holds model defaults in [models.new_thread].
A named AWS profile for CLI sign-in; Claude Desktop can use native IAM Identity Center sign-in.
A Terraform module deploys account guardrails in two states: the management account holds the SCP, and each child account holds its profiles, trail and alarms. StackSets bootstrap deployment roles across an OU; Terraform deploys the resources into each child account.
Application inference profiles receive token anomaly alarms, and team tags support cost allocation. Runtime OpenAI usage is attributed by role session.
One per approved model, tagged per team.
Attached to the target accounts or OUs.
Multi-region Bedrock events to encrypted, versioned S3 with a configured 30-day lifecycle.
Input and output token usage per inference profile.
Two client families, one config.json, a named AWS profile, and shared account guardrails.
Claude Code availableModels limits picker aliases. Codex requirements.toml enforces the provider and connection policy. Neither is a model authorization boundary: use IAM and SCPs.
| Control | Where it lives | Boundary? | What it does |
|---|---|---|---|
| Claude Code availableModels | Client managed settings | No | Limits the aliases in the model picker, not raw model IDs |
| requirements.toml | Managed client policy | No | Enforces the Bedrock provider and AWS connection settings; model defaults are changeable, not a model allowlist |
| IAM permission set | AWS IAM Identity Center | Yes | Authorizes which Bedrock models and profiles a person may invoke |
| Service control policy | AWS Organizations | Yes | Denies direct foundation-model calls, with temporary exceptions for Global OpenAI system profiles and the default project; IAM still authorizes access |
| CloudTrail and anomaly alarms | AWS account | Audit | Records who called what and alerts on unusual token usage; does not block calls |
Start with AWS access, deliver managed configuration, then verify both successful calls and denied requests.
Enable IAM Identity Center, create a Bedrock permission set and deploy bedrock-governance.
Fill in one config.json in coding-agent-3p and render the files for each client and OS.
Deliver macOS profiles through MDM. On Windows, deploy Claude policy and Codex requirements.toml to their managed locations.
Run the test plan: sign-in, model calls, the SCP denial and the CloudTrail record.
ASCENDING uses automation and Infrastructure as Code to spin up your entire Bedrock governance framework within 3 business days - no lengthy professional services engagement, no manual configuration, no disruption to your teams in the process.
Our automated IaC templates deploy the complete governance framework into your AWS environment - wiring up your identity system, configuring team access policies, and activating cost attribution so every team's spend is visible from hour one.
We review your internal data sources and current application integrations, validate your security architecture to ensure all AI workloads stay inside your AWS boundary, and deploy the spending monitoring agent that watches your budget continuously.
Your operations team receives everything needed to manage the environment independently - runbooks, escalation paths, and a clear understanding of how to adjust policies and budgets as your AI usage evolves over time.
Use the guide for the client your team runs, with the same account-level governance foundation.
Enterprise sign-in, managed desktop policies, model access and optional Jarvis governance.
Managed Runtime configuration, AWS SSO, model boundaries and deployment validation.
Common questions about client configuration, account guardrails and attribution.
3P means the client sends inference to a third-party platform instead of the vendor’s hosted service. On Amazon Bedrock, Claude Desktop and Claude Code call models in your AWS account, and people sign in through AWS IAM Identity Center.
Yes. Managed configuration points them at the built-in amazon-bedrock-runtime provider and refreshes AWS SSO credentials; no translation gateway is needed.
No. Claude Code availableModels limits aliases, not raw model IDs, and requirements.toml enforces provider and connection settings, not a model allowlist. Use IAM permissions or an SCP as the model boundary.
For Claude, requiring application inference profiles keeps usage on named, tagged resources. The current module has a temporary exception for Global OpenAI system profiles and the default project; attribute ChatGPT and Codex usage by role session.
As two Terraform states: the management account holds the SCP, and each child account holds its profiles, trail and alarms. StackSets bootstrap deployment roles across an OU; Terraform deploys the resources into each child account.
No. Clients call Bedrock directly with their own AWS credentials. Jarvis is an optional layer for additional policy and attribution on traffic routed through it.
For Claude, application inference profiles tagged per team carry cost and token metrics. For ChatGPT Desktop and Codex, attribute usage through the person’s role session in CloudTrail or Bedrock invocation logs.
The coding-agent-3p and bedrock-governance repositories are open source and free. ASCENDING delivers implementation, configuration and support as a paid engagement.
ASCENDING helps your team deploy managed configuration, AWS sign-in and account guardrails.
