Found a cool thing for anyone who works a lot in Codex, Claude Code, Gemini and other AI tools
It's called CodeBurn
You can see:
- which model eats the most
- which project turned out to be the most gluttonous
- how many tokens went to different types of tasks
- which sessions turned out to be the most expensive
- how much you spent today / this week / this month
The handiest part: CodeBurn pulls data from the local history of your AI tools. No need to connect API keys or do anything complicated
So you install it - run it - and it can already break down your past sessions
🧩HOW TO TRY IT
Repository: https://github.com/getagentseal/codeburn
You don't even have to deal with the installation by hand
Open Codex/Claude and send:
Study the README of the GitHub repository https://github.com/getagentseal/codeburnCheck the requirements and the installation method in the author's current instructions
Install CodeBurn in a safe way and launch its web interface
Don't change anything in my current project
After launching, explain to me in simple words:
- where to see token usage;
- which projects and models use the most;
- which types of tasks the tokens go to;
- where the most expensive sessions are shown;
- what looks inefficient in my usageIf anything needed for installation or launch is missing on the computer, first explain what exactly is needed, and only then install it
After that a dashboard with stats opens up, and you can calmly dig into where the spend goes
By the way, a handy thing after our material about limits: here you can already look at YOUR OWN data and see where exactly you're burning them
⚠️ Small catch: if Codex or Claude runs on a subscription for you, the dollar figure you see is a rough equivalent at API prices, not real money charged from your pocket, so don't freak out)) But you can estimate how much you spend on the subscription and how much it would cost you if you paid for the API
Test it out. I think this thing is seriously super useful and lets you understand where your limits go
How to optimize? Just ask, it'll give you pinpoint recommendations
Why limits vanish without you noticing
Usually it's not one huge task but small things repeated hundreds of times. A long session drags the whole conversation history along, and every new question is paid for together with it. A strong model gets called for simple work a fast one could handle. A scheduled agent wakes up and rereads the same things even when there's nothing new. CodeBurn is good because it shows this on your own data, not in theory
What helped me
First and most important, roles for models. I have 8 Claude Code agents working every day, with different models: Opus 5.5 where depth and architecture are needed, Sonnet 5 where speed and volume matter more. The expensive model isn't spent on work a fast one can do
Second, don't make the model do what a regular program can. In the content factory a bot reads 13 sources around the clock, and duplicates are dropped, so only the best topics reach a draft. The less extra stuff reaches the model, the fewer limits you burn
Third, short sessions for one task. Finish a job, start the next one from a clean slate instead of dragging the whole day's history along. The model doesn't reread old talk on every question and answers more precisely
How to read the numbers if you're on a subscription
The dollar amount in dashboards like this is useful for comparison, not panic. It shows what your work would cost if you paid per token, and helps you decide whether the subscription is the better deal. For a small team a subscription almost always wins, and the dashboard helps you see where exactly you hit the limit
If you want someone from the outside to look at where your AI processes overspend, I do that in an AI audit