Agent blowout calculator

Embed it on your site

Two lines. No framework integration, no build step, no API key. Drop the mount div where you want the widget, then load the script once anywhere on the page:

<div id="agent-13-embed"></div>
<script async src="https://keybrake.com/embed.js"></script>

The widget is vanilla JavaScript with inline styles — it doesn't pull a stylesheet, doesn't read cookies, doesn't set any, and doesn't fire an analytics beacon. It ships a "Powered by Keybrake" attribution link; keeping that link is the license.

Works on

• Static sites (Jekyll, Hugo, Astro, 11ty, plain HTML)
• React / Next.js / Remix — use dangerouslySetInnerHTML for the mount div, or render the <div> normally and add the script in a <Script strategy="afterInteractive">
• Vue / Nuxt, Svelte / SvelteKit, Solid — same pattern
• Webflow, Ghost, WordPress — paste into a custom HTML / embed block
• Substack, Medium — not supported (they strip script tags)

How the math works

The calculator does one thing: multiply a stuck-loop rate by a per-call cost, then compare against a daily cap. The numbers you can tune are the calls-per-minute slider (1 to 600) and the vendor; the per-call cost comes from a short table baked into the widget:

The Stripe number looks absurd, and it is — that's the point. A stuck agent hitting Stripe at one charge per second with a mid-market average ticket would burn $1.3M before the day is out. Real-world incidents top out lower because agents don't sustain that rate for 24 hours, but a single bad retry loop at $15 / charge for five minutes is $4,500 of charges you have to reverse. Per-vendor caps don't eliminate the risk; they bound it.

For the "time to burn $100" readout: we invert the rate. At 60 Stripe charges per minute (one per second, $15 each), $100 takes 6.7 seconds. At 60 Twilio SMS per minute that's 3.5 hours. At 60 Resend emails per minute, almost three days. Cheap calls are not the blowout surface — cost-per-call × rate is.

What the calculator is not

It's not a real-time burn monitor. It doesn't know your actual rate or actual vendor spend; it shows what could happen given the inputs. For real enforcement you need a proxy between your agent and the vendor — one that parses the cost out of each response and blocks the next call when the cumulative total crosses your cap. That's what Keybrake does. The calculator is the visualization of that policy loop; the proxy is the loop itself.

It's also not prescriptive about what your cap should be. $50 / day on Stripe is a starter default for an agent that processes a handful of real charges per day. If your agent does one charge per week with an average ticket of $2,000, your cap might be $2,500 / day (one-and-change charges) and an alert at $5,000. The calculator nudges you to think about the ceiling explicitly; picking the ceiling is your call.

Background reading

• AI agent kill switch: 4 patterns, measured in seconds — the caps in the widget are pattern 3 (cost-circuit-breaker), with measured stop-latency numbers.
• The 2026 agent governance stack: which proxy goes where — where SaaS-API caps sit in the broader four-layer governance stack.
• How to give an AI agent a Stripe API key without setting your money on fire — the Stripe-specific practical guide, with the minimum restricted-key scope.
• Keybrake vs LiteLLM — why this tool measures USD parsed from response bodies, not tokens.