Codex usage tracking on Mac should not wait for the invoice

If you use Codex, Claude Code, Cursor, ChatGPT, or the OpenAI API for real work, token usage has a weird visibility problem.

You usually find out what happened after the session is over.

That is fine for a receipt. It is bad for a tool you are actively steering.

The problem with after-the-fact AI usage

AI coding work is full of loops that feel small in the moment:

• retrying the same failing task
• pasting a huge file when a smaller excerpt would work
• asking an agent to scan more context than it needs
• switching models without noticing the cost difference
• letting a session run because the output is almost right

None of those moments feels like a billing event.

But together, they are exactly how AI spend creeps up.

By the time a dashboard or invoice shows the total, the useful decision point has already passed.

A token counter belongs where the work is happening

For Mac users, the best place for AI usage visibility is not another analytics tab. It is the menu bar.

A menu bar token counter gives you a small ambient signal while you are still working:

• how much context this session is burning
• whether a Codex or Claude run is getting expensive
• whether retries are starting to stack up
• when a prompt should be trimmed before sending
• whether usage is normal or drifting into bill shock territory

That is the difference between usage reporting and usage control.

What I would track during Codex or Claude work

For AI coding sessions, the useful numbers are not complicated:

1. Tokens used in the current session
2. Approximate cost before the bill arrives
3. Provider and model visibility
4. Usage over time, not just at the end
5. A fast way to notice runaway context or retry loops

That is enough to change behavior.

If the number is visible, you make smaller prompts. You split tasks earlier. You stop asking an agent to reread the whole repo when one file would do.

TokenBar

I built TokenBar for this exact reason: a Mac menu bar app for seeing AI token, cost, and usage signals while you work.

It is aimed at people using tools like Codex, Claude, OpenAI, and other AI workflows where spend can hide inside long sessions.

Link: https://tokenbar.site/

The goal is simple: know what your AI work is costing before it turns into bill shock.

The practical takeaway

If you use AI coding tools every day, do not treat token usage like a monthly receipt.

Treat it like a live instrument.

The earlier you can see usage drift, the easier it is to fix the prompt, switch models, stop a retry loop, or break the task into something cheaper.