When I use Claude Code, Codex, Cursor, or any other coding assistant, the number I care about is usually not the total token count.
The useful question is simpler:
Is this session still producing value, or am I just paying for the model to circle the same problem?
That is the pattern that pushed me to build TokenBar, a small macOS menu bar app for watching AI usage while I work.
Site: https://tokenbar.site/
Most AI coding tools show you the magic part. They generate files, suggest patches, explain errors, and keep going until the task feels done.
The cost and context side is much quieter.
That makes it easy to miss a bad session until it is already obvious in hindsight.
The problem is not one expensive request
One big request is usually fine if it moves the project forward.
The waste shows up as a loop:
- Ask the agent to fix something
- It reads a lot of context
- It changes the wrong layer
- You correct it
- It reads the context again
- It repeats the same assumption
At that point the assistant is not helping less because it is bad. It is helping less because the session has lost shape.
The token count becomes a signal that the workflow needs a reset.
Session usage beats monthly usage
Monthly usage is useful for billing.
Session usage is useful while you are still making decisions.
If I can see that one debugging pass is getting weirdly expensive, I can change how I work right now:
- restart the agent with a tighter prompt
- paste only the relevant files
- split the task into a smaller bug
- stop asking for broad refactors
- switch from agent mode to manual editing
That is the difference between a receipt and a dashboard.
A receipt tells you what happened.
A live session signal changes the next move.
Good AI coding has a budget shape
The best sessions usually have a shape:
- a small discovery phase
- one or two focused implementation passes
- a test or verification pass
- a short cleanup pass
The bad sessions have no shape. They just keep consuming context.
That is why I do not think token visibility is only about saving money. It is also about noticing when the collaboration is getting sloppy.
If a human teammate asked for the same context five times and kept editing the wrong file, you would pause and reframe the task.
AI coding should get the same treatment.
Why put this in the menu bar?
I did not want another analytics page.
When I am coding, I want a tiny signal that stays visible without becoming another app to manage. The menu bar is boring, but that is the point.
A small number in the corner is enough to make me ask:
Is this still a good use of the model?
That little interruption can save a surprising amount of wasted work.
The lesson for AI tool builders
If you are building around coding agents, I think there is a useful product lesson here.
Developers do not only need more automation.
They need better signals for when automation is drifting.
The most useful AI tooling is not always the tool that does more. Sometimes it is the tool that helps you stop sooner, restart cleaner, or notice that the current path is getting expensive.
That is the lane I am exploring with TokenBar: lightweight usage visibility for AI coding sessions, sitting in the Mac menu bar instead of buried in a billing page.
Top comments (0)