noburn

noburn is a guard that sits in front of your LLM calls. Before each request leaves your process, it runs a check against a budget and returns allow or refuse with the estimated cost and remaining balance — so a call that would blow the cap is never sent to the provider and never billed, rather than being caught in next month's invoice.

The problem it names is agent-specific: agents do not fail by falling over, they keep working correctly at a cost nobody is watching. The site walks through the shapes that do it — a retry loop that runs all night, context that compounds so turn forty costs twenty times turn one, and a runaway batch — and offers three limits checked on one request: per-user, per-run and per-project.

It ships as two SDKs, `@noburn/sdk` on npm and `noburn` on PyPI, both listed at version 0.1.0 under MIT with zero additional dependencies, plus a `bench-check.mjs` script in the repo for measuring the added latency against your own project. The site is explicit that it is pre-launch and has no customer logos yet, and publishes a worked example of a refused call.

Capping spend per user, per run and per project; stopping an infinite retry loop from billing all night; guarding agent runs with compounding context; adding a budget check to an existing LLM integration.

Official privacy policy