Quick answer
Per-developer Codex usage comes from two sources. Local session files show tokens, models, and projects on each machine. Organization reporting shows seats and billing. Join them on the developer identity, and treat any number that comes from only one source as partial.
Two sources, two questions
| Question | Source | Granularity |
|---|---|---|
| Who used the most tokens? | Local session files | Per machine |
| Who has a seat? | Admin console | Per user |
| What did we pay? | Billing | Per organization |
| Which project is expensive? | Local working directory | Per repository |
Collect without surveillance
Session files contain prompts and code. Collecting them centrally is a privacy decision, not just a reporting one. Prefer aggregate exports over raw file collection.
- Agree which fields are collected: tokens, model, project, date.
- Export aggregates, not transcripts.
- Set a retention period and delete raw exports.
- Report by team or project before reporting by person.
Use the numbers well
- Do not rank developers by tokens; rank workflows by cost per shipped change.
- Compare the same task type across teams.
- Watch for one project consuming a disproportionate share.
- Use the data to buy the right plan tier, not to police typing.
Reporting team usage without surveillance
Per-person token counts invite the wrong conclusions: a large number can mean hard work, a runaway agent, or a task that should have been automated. Report by project and task type first, and use per-person numbers only to size seats.
- Agree which fields are collected: tokens, model, project, date.
- Export aggregates rather than prompts.
- Set a retention period and say what it is.
- Review team totals before individual totals.
- Give each developer their own numbers first.
The AI coding cost tracker guide covers the money side, and Codex limits by plan covers seat sizing.
Questions, answered
Does Codex offer an admin usage dashboard?
Business and Enterprise plans include organization reporting for seats and billing. It does not break tokens down by local project.
Can UsageAtlas report on a whole team?
It runs per machine. Teams combine exports or use the optional cloud sync, subject to their own privacy rules.
Should we collect session files centrally?
Usually not. They contain source code and prompts. Collect aggregate token records instead.