Claude Code, Cursor, Copilot, four API keys — and no idea what your AI costs

Vibe coding in 2026 looks like this: Claude Code open in one terminal, Cursor humming in another, a Copilot tab you forgot you were logged into, and somewhere in a password manager, four API keys you spun up for “just one experiment” and never thought about again. Four providers. Four meters running. And your actual AI bill at the end of the month? A complete mystery.

I live inside that mystery, because I build this way every day. So I made a thing to answer one question — what is AI actually costing me? — and I made it free. No card, no paid tier, no upsell, ever. It’s called MyTokenTracker.

Why this even needs to exist

If you only ever called one model from one SDK, you wouldn’t need me — you’d read the provider dashboard and move on. But that’s not how anyone actually works now. The modern developer’s spend is smeared across Claude Code, Cursor, Copilot, a couple of agent loops, and three raw API keys, and every one of those bills lives in a different tab, in a different format, on a different billing cycle.

Vibe coders have it worst of all. When you’re generating whole files at a time across three tools, the tokens move faster than your intuition does. The spend is invisible right up until the invoice lands — and by then it’s a single scary number with no story attached. Which model ate the budget? Was it the reasoning tokens? The context you kept re-sending? The one agent that silently retried forty times? You can’t answer any of that from a total. You need the breakdown, in one place, for all of it at once. That’s the hole this fills.

Wire it up in one line

The whole point is that it can’t cost you effort either. If you’re in Claude Code, it’s a single line:

curl -fsSL "https://mytokentracker.io/install.sh?token=YOUR_TOKEN" | bash

From there it just watches. Every session, every model, captured automatically — you go back to building and the dashboard fills in behind you. Not on Claude Code? There are drop-in Python and Node wrappers for OpenAI, Anthropic, Gemini, and Mistral — swap one import and you’re tracked. And if your stack is something weirder, one POST to the events API covers literally any provider you can make an HTTP call to.

What it actually tracks

The geeky part — this is the stuff I built it to see:

• Cost by provider, model, platform, and use-case — so “my AI bill” stops being one number and becomes a list of decisions you can actually argue with.
• The tokens nobody bothers to count — cache reads and reasoning tokens, including the Gemini “thoughts” that Google folds into output billing. If you’re paying for it, you should be able to see it.
• The ratios that tell the real story — tokens per dollar, latency, success rate. Cheap-but-failing is not actually cheap.
• Real cost when it knows it, an honest estimate when it doesn’t — the estimates come from a price catalog that re-syncs every day across 2,300-plus models.

And because a bill only matters if you can see where it’s heading, there’s a part I’m a little too proud of: an AI Cost Index — think an S&P 500 for token prices — live pricing for those 2,300-plus models, and a real-time “State of AI” page that pulls the whole moving picture into one place.

“Free forever” — and I know how that sounds

Here’s the part a developer always wants answered before they pipe a script into bash. “Free forever, open data, no catch” makes the back of my neck itch too — because most of the time, if something sounds too good to be true, it is. The free tool is usually the funnel; the data is usually the product, and you’re it.

So here’s the honest answer to “what’s the catch.” The data is open under CC BY 4.0 — JSON, CSV, an API, yours to take. The tracking client is MIT-licensed and on GitHub, so you can read every line that touches your usage before you run that install command. No account wall, no “contact sales,” no premium tier dangling the one feature you need. The catch, if you insist on one, is that this isn’t a funnel — it’s a research tool. Track once, and the shared picture of what AI really costs gets a little sharper for all of us. That’s the whole trade: I get a clearer map, and so do you.

And you don’t have to take my word for any of it — the license is public, the code is public, the data is public. That’s the entire reason I built it in the open. Go kick the tires.

See what you’re really spending →

It’s free, it always will be, and it runs from your keys to your own dashboard. Wire it up once and stop guessing at the number — mytokentracker.io.

(And yes — the little guy on the State of AI page is our mascot. He’s free too. No catch on him either.)