AWS Reveals How to Track Codex Usage with CloudWatch

Summary

The AWS ML blog introduced an architecture that routes OpenTelemetry metrics from Codex running on Amazon Bedrock to CloudWatch
A local collector on the developer's workstation adds organizational context to the metrics, signs them with AWS SigV4, and sends them to the CloudWatch OTLP endpoint
Without a central proxy or separate infrastructure, adoption status and cost-related signals can be viewed by user, team, department, and cost center

AWS has revealed how to connect the OpenTelemetry metrics generated when using the coding agent Codex on Amazon Bedrock to Amazon CloudWatch. When a developer signs in via IAM Identity Center and runs Codex with Bedrock as the model provider, a collector listening on local 127.0.0.1 receives the metrics, adds organizational context, signs them with SigV4 using temporary credentials, and sends them to the CloudWatch OTLP endpoint.

A notable feature of this approach is that it does not add a central proxy to the model request path. The reference deployment creates only a CloudWatch dashboard, without an ECS service, load balancer, VPC, or public ingestion endpoint.

What You Can See

The built-in CodexOnBedrock dashboard shows 24-hour rolling totals for active users, conversation turns, API requests, and token usage. Views are also provided by model, token type, user, department, team, cost center, organization, and session source, which can be used to identify how far adoption has spread and where consumption is concentrated.

AWS emphasized that token counts are not a billing ledger. Since actual costs can vary depending on pricing changes, discounts, and credits, the company explained that understanding real spend requires using AWS Cost and Usage Reports (CUR) 2.0, which provides IAM-principal-based cost allocation, or Bedrock cost management reports.