Last month I got a $340 bill from Anthropic. I knew I'd been using Claude a lot but had no idea it had gotten that high. The usage dashboard is fine but you only check it after the damage is done.
So I built something that shows the cost per call in your VS Code status bar as it happens. It runs a local proxy, you point your SDK at it with one env var, no code changes. Nothing leaves your machine.
Would love to know if anyone else has had this problem or found a better way to deal with it.
https://marketplace.visualstudio.com/items?itemName=tokenlens.tokenpulse
Seeing cost in the moment instead of after the invoice is the right move. Once AI usage is in the edit loop, people will overspend unless the signal shows up before they hit enter. If you keep building this out, I'd add soft budget caps and attribution by feature or workspace, because "one Claude call" stops being useful the second a team shares an API key. We ran into the same pattern building DictaFlow: usage gets a lot easier to manage once the feedback is inside the workflow instead of buried in a dashboard.
A $340 surprise bill is exactly the kind of thing that makes people actually go check where the tokens went, so good timing on building this. When I run multiple coding agents in parallel the cost tracking rabbit hole is usually less about the sticker price and more about figuring out which specific agent run or retry loop burned the budget, sometimes one bad prompt spirals into ten retries before anyone notices. Does your extension break usage down per session or per file being edited, or is it more of a running total across the whole workspace?
Yeah this is a real pain point—cost visibility usually comes too late, especially when usage is spread across dev/testing. Catching spend per call in real time inside the workflow itself is exactly the kind of thing that should be default but isn’t.
This is useful because it moves cost visibility into the decision loop, not after the invoice arrives.
One thing I would add if this ever moves from “developer tool” to “user-facing AI app” is request-level attribution: provider/model, user or workspace, feature, idempotency key, and success/failure state.
I am working on SettleMesh, and the pattern we keep seeing is that once an AI action can spend money, billing starts to look more like a ledger than a subscription page. Your status-bar view feels like a great first half of that story: make the cost visible before it turns into a surprise bill.
I like that you're making cost visible at the moment decisions are made instead of after the invoice arrives.
A lot of developer tools report what happened. Showing the cost while someone is still writing code changes behavior, because developers can adjust before the expense becomes irreversible.
This comment was deleted a month ago