A bill you cannot refuse is a bug.

The problem

Reporting is not control.

An LLM call is billed the moment it is answered. By the time a dashboard has a number to show you, the money is gone. That is fine when a request costs a fraction of a cent and a human is waiting on it. It stops being fine when an agent loops, a retry storm compounds, or one customer’s workload runs unattended overnight.

The only place a spend limit can be enforced is before the request is sent. So that is where noburn puts it: a check that returns allow or block, against project, per-end-user and per-run caps, in front of the provider you already use. A blocked call is never billed, because it is never made.

How that differs from gateways and tracing tools.

Who writes this

Orvi.

I build and maintain noburn, and I write everything published on this site. The posts name specific tools, quote specific prices and describe specific failures, so they carry a name rather than a masthead: if something here is wrong, it is wrong because I got it wrong, and I would like to hear about it.

Comparisons on this site name competitors directly and cite vendor documentation with the date it was checked. Where a competing tool does something well, it says so. That is the standard I would want applied to noburn.

Corrections and questions: open an issue.