Set up OpenTelemetry for OpenAI Codex
OpenAI Codex emits OpenTelemetry (OTel) metrics about token usage, per-turn latency, API requests, tool calls, and approvals. When you send these metrics to Amazon CloudWatch, they populate the Codex tab of the Coding Agent Insights dashboards. For more information about the dashboards, see Coding Agent Insights.
Codex configures its metrics exporter through the [otel] section of its
config.toml file. Two setup paths are available, depending on how your
developers run Codex.
Choose a setup path
-
Bearer token — Codex sends OTLP metrics directly to the CloudWatch native OTLP endpoint, authenticating with an
Authorization: Bearerheader. Identity and organizational attributes are supplied through theOTEL_RESOURCE_ATTRIBUTESenvironment variable. This path needs no collector and is the fastest way for an individual developer or a small team to get started. See Set up OpenAI Codex with a bearer token. -
Enterprise rollout — For organizations that federate developer identity through corporate single sign-on (SSO) such as Okta, Microsoft Entra ID, Auth0, or AWS IAM Identity Center, centralize Amazon Bedrock access, and distribute a standard configuration to a fleet of developers. Identity and organizational attributes are derived from your identity provider (IdP). See Set up OpenAI Codex for an enterprise rollout.
Note
The bearer token path configures only how Codex sends
telemetry to CloudWatch; it is independent of how Codex calls a model
for inference, so you can use it with any model provider. The
enterprise rollout, in contrast, is built around centralized Amazon Bedrock access, so
Amazon Bedrock is the inference backend for that path. To use Amazon Bedrock for inference with
Codex, see Amazon
Bedrock
For the complete, continually updated guidance, including deployment templates and
setup scripts, see the guidance for Codex on Amazon Bedrock
