I accidently built a tool that devs needed way more than me.

This started as a small itch.

My LLM provider (OpenAI, to be exact) gave me one number: total spend, aka a bill. But I wanted to know which function in my app spent more, and why. The code or the summarizer? Obviously the code, right?

There was no way for me to see my usage broken down. So I built a tiny thing for myself: one line after each API call, sending the token counts and a feature name. I only wanted visibility, nothing more.

Then I got curious whether anyone else cared, and went looking. My problem turned out to be the door into a rabbit hole.

• A developer on Google's AI forum found about €290 in search charges attached to roughly €10 of actual LLM usage. Their spend cap never caught it.

• Another dev wrote about a job that blew up their bill while every dashboard stayed green. A timeout plus an automatic retry was quietly doubling costs on long-context calls.

• At the far end, one startup reported an AI bill of $113,000 in a single month.

For me, the missing breakdown was curiosity. For them, it was money leaving before anyone could see where it was going.

So I went on to build more features and test with more functions. Every guide on writing a good blog post says I should mention my expertise so readers will trust me, so here it is: I have little to no expertise in building products. But I do have experience solving other people's problems, so I went on researching, adding more functions, testing with more providers and models, learning words like proxies and gateways, and studying companies like Langfuse.

The tool I built for my own visibility answers exactly their question: which feature, which model, which customer, the second a request completes. So I kept building. It's now llm-track.com. No proxy, and prompts never leave your servers.

I'm documenting the journey from here. And I bet you already guessed the answer: of course it was the code, spending most of my money.