Claude was given one plain-English task and one tool. It decided on its own to call that tool — which happened to be a live, x402-gated HTTP endpoint. That decision autonomously settled a real USDC payment on Base mainnet, mid-conversation, before the agent could see the data it asked for.
Task, not instructions. The agent is told what a trading desk needs — a liquidation-risk briefing — not how to fetch it.
Agent decides to pay. It calls the one tool it has. That tool is a real x402 endpoint — the 402 challenge and USDC settlement happen underneath, invisibly to the model.
Real data, real answer. The paid response flows back into the same conversation. The agent reasons over live data and writes the briefing.
That is the whole thing an agent framework needs to know. No SDK, no wallet plumbing exposed to the model — just a tool description an LLM already knows how to call.
Built on the x402 protocol. Two live products right now: liquidation-risk scoring and a 25-endpoint toolbox (oracles, escrow, invoicing, subscriptions).