Hey PH 👋 (I'm Tyler, maker of Burnwatch.)

A year ago, "my AI agent spent money" was a thought experiment. Today agents pay for APIs, search, and compute on their own, over x402 and stablecoins. That's the exciting part. The terrifying part is that an agent has zero instinct for "wait, this looks wrong."

A bug, a runaway loop, or a prompt injection doesn't crash. It just keeps paying. By the time you glance at Slack, the wallet is empty.

We already solved this for people. When your credit card does something out of character, your bank pings you in seconds. Agents have had nothing like that.

Burnwatch is that alert, for your agents.

You add one line after each payment (or wrap your x402 client). Burnwatch learns each agent's normal, then fires the moment it drifts: burn rate spiking, a counterparty it's never paid, draining in bursts, spending at 3am, jumping to a brand new rail. Every alert ships with the evidence (the exact numbers and rules that tripped), so you can tell a real incident from noise in one glance.

Two things I refused to compromise on:

→ It never holds your keys and never sits in the payment path. Metadata only. If Burnwatch goes down, your agent keeps paying. Monitoring should never be able to break the thing it's monitoring.

→ It's trivial to try. No wallet required. "pip install burnwatch", or open the live demo at app.burnwatch.dev/live and watch an agent get hijacked in real time, detection firing in under two seconds.

The free tier is genuinely usable: 2 agents, 10k events a month, every detection rule included.

If you're building agents that spend, I'd love to hear: what would you want flagged that nothing catches today?