A few months ago I had one of those dumb founder moments where you stare at a bill and think: wait, how did it get that high?
Nothing was obviously broken. No outage, no exploit, no wild traffic spike. I was just using LLMs all day while building, testing prompts, comparing models, and bouncing between apps.
The problem was simpler: I had no feel for token burn while I was working.
I could check dashboards after the fact, but that is like checking your speedometer after the road trip is over. Useful for accounting, useless for control.
So I built TokenBar.
What was actually missing
Most AI tooling gives you usage data in one of two bad forms:
- buried in provider dashboards
- surfaced only inside one app or one playground
That is fine if you are doing occasional experiments.
It is not fine if your whole day is split across ChatGPT, Claude, Cursor, API playgrounds, and your own product.
I wanted one thing:
a live token counter sitting in the macOS menu bar so I could see usage while I worked
Not later.
Not in another tab.
Not after I already made ten expensive calls.
Just a running sense of: am I being careful right now, or am I lighting money on fire?
What changed once I could see it
The biggest win was not optimization in the abstract. It was behavior.
Once usage became visible, I started catching patterns immediately:
- retrying prompts that were too vague
- using a large model when a smaller one was enough
- letting long context accumulate for no reason
- forgetting that a background workflow was making repeated calls
None of this is dramatic on a single request.
Over a week, it adds up fast.
That is the annoying part about AI spend. It often feels cheap in the moment and expensive in the aggregate.
The founder lesson
I keep relearning the same lesson as a solo dev:
if a cost is invisible, it will drift upward
It is true for cloud infra.
It is true for churn.
It is true for AI usage.
Dashboards are great for reporting.
They are not great for moment to moment decisions.
Visibility changes behavior faster than advice does.
Why I think this matters now
A lot of builders are adding AI features before they have a strong instinct for token economics.
That makes sense because shipping speed matters. You do not want to spend three weeks building internal cost tooling before validating demand.
But once usage grows even a bit, the blind spot gets expensive.
You do not need a giant finance stack.
You just need tighter feedback loops.
That is what I wanted from TokenBar: a tiny piece of ambient awareness for AI usage, the same way battery percentage or Wi Fi status sits in the corner of your screen.
If you are building with LLMs
My suggestion is simple:
track token usage close to where decisions happen
If that is a menu bar app, great.
If that is logs in your own app, also great.
If that is a custom internal dashboard, fine.
Just do not wait until the monthly bill teaches the lesson for you.
I built TokenBar because I got tired of learning that lesson the expensive way.