Put a certified cost & risk bound on your AI agents.
costwright measures your agents and emits a signed certificate with a conservative bound on their cost and risk — a verifiable ceiling backed by statistics, recomputable by a third party. Part of the bound is mechanized and verified in Lean. Not an estimate, not a dashboard.
Open-source library · free tier · no credit card.
An estimate is not a guarantee.
Every cost tracker shows you a number after the fact. None of them give you the ceiling that holds before you ship. When an agent loops or fans out, the spend you forecast and the spend you get are not the same — and you find out on the invoice.
- ▸ An average run cost tells you nothing about the worst case you'll actually pay.
- ▸ Observability tools (dashboards, token counters) report spend — they don't bound it.
- ▸ Nobody else ships a signed, recomputable cost & risk bound, with a Lean-verified core.
$ costwright certify --agent my_agent --runs runs.jsonl modeling cost & risk over runs … fitting conservative bound (Lean-verified core) … ✓ certificate issued cost ≤ $0.0042 / run (95%, n=1200) risk ≤ 0.7% (95%) bound is conservative by construction signed · recomputable · input-hash bound
How it works
Measure
Point costwright at your agent's runs (token spend, tool calls, samples). It models the cost and risk distribution over independent runs — the real spread, not a point estimate.
Certify
costwright fits a conservative bound — conservative by construction, it never understates the cost or risk — and emits a signed certificate binding the bound to the exact inputs (hash). Recomputable by anyone, not a number you have to trust.
Guarantee
Ship the certificate as your ceiling: cost ≤ X at a stated confidence. The core bound is mechanized and verified in Lean (a formal proof), so the guarantee holds — auditable by a third party, not a vibe.
The certificate is the product.
A verifiable, recomputable certificate that a hollow product could not emit. Every certificate carries the bound — cost, risk, confidence, input hash — not a sensation. The signal sells without a salesperson.
Pricing
Start with the open-source library. Pay when you need the hosted certification API or a signed report.
- Starter · Pro · Scale
- Certify cost & risk at scale
- Founding access — early teams
- Signed PDF + bound + input hashes
- Cost & risk ceiling, Lean-verified core
- For a regulated review
Questions
Is this another cost dashboard?
No. A dashboard reports what you already spent; costwright bounds what you can spend. We emit a ceiling — cost ≤ X at a stated confidence — conservative by construction, not a number you watch after the fact.
Why should I trust the bound?
Because part of it is mechanized and verified in Lean — a formal proof — and the whole bound is conservative by construction (it never understates cost or risk). The certificate is recomputable, so a third party can check it independently.
Does it work on my agents?
Yes — costwright wraps any AI agent or pipeline. You bring the runs (cost, tool calls, samples); it models them and certifies a conservative ceiling. The base library is open-source.
Founding access?
A founder price for the hosted certification API while we onboard the first teams. The library stays free and open-source.