Skip to main content

​
Codex CLI

OpenAI Codex CLI ships a built-in OpenTelemetry exporter. Point it at MeshAI’s OTLP logs endpoint and every Codex session is monitored with no package to install: per-turn token usage and cost, tool approval decisions, sandbox outcomes, and session activity.

​
Configure

Add an [otel] block to ~/.codex/config.toml:
Use an API key with the telemetry:write scope. The next Codex session (interactive or codex exec) starts exporting; the agent appears in your MeshAI registry within about a minute as codex_exec or codex_tui, framework codex, with each Codex conversation tracked as a session.
The API key sits in plaintext in config.toml, so keep the file chmod 600 and use a key scoped to telemetry:write only. Codex has no environment-variable indirection for exporter headers yet.

​
What’s tracked

​
Privacy

Only usage, model, and session metadata is persisted. MeshAI’s ingest maps Codex events onto OpenTelemetry GenAI semantic conventions at the boundary and drops everything else: prompts (already [REDACTED] by Codex unless you opt in), user.email, and auth diagnostics never reach storage. Leave log_user_prompt = false; the Codex log integration does not use prompt text.

​
Durability caveat

Codex’s exporter is in-process and fire-and-forget: events buffered when the process crashes or the network is down are lost. That is fine for monitoring and cost attribution. If you need evidence-grade capture with a local write-ahead log, that is what the Claude Code connector provides for Claude Code; a durable Codex connector is on the roadmap.

​
Requirements

Codex CLI 0.146 or later (earlier versions with the [otel] config section also work; telemetry verified against 0.147).