Watch an agent get stopped.
$ npx @thsky-21/thskyshield-demoAn agent loops, overspends, and gets stopped — about four seconds, no signup, no API key.
The whole run, exactly as it prints.
$ npx @thsky-21/thskyshield-demoA real capture of the published package, not a mock-up. The model call is simulated so the demo needs no LLM key, but the prices come from the same registry this product bills against and the decision around each call is the real one. Your own run will differ in one place only: the milliseconds.
What the four seconds prove.
Cost is tracked per step
Every step is priced before it runs — $0.003250 for 500 in / 200 out on gpt-4o, from the same registry that prices production runs. You watch the total climb line by line, from $0.003250 at step 01 to $0.048750 by step 15.
The loop is caught by prompt fingerprint
The prompt is hashed locally — the text itself is never stored or sent. The same fingerprint comes back every step, and the counter climbs from ×2 at step 02 to ×15 at step 15, against a gate of 20. It never reaches the gate: the money runs out first, which is what a real runaway usually looks like.
The run is stopped before the next call fires
Step 16 was denied before the model was called — not logged after the fact. The run ended at $0.048750 against a $0.05 ceiling. Under it, not over it, because the step that would have crossed the line never ran.
Now point it at your own agents.
That was a local simulation. Point it at your own agents and the same engine runs in observe mode — it records what it would do and totals what that would have saved, without touching your runs.
Already ran it? Wrap your own loop in about five minutes →