AisleRun：AI agent 付费调用的结账测试服务
为 MoonPay PayBox 黑客松做的按需结账测试服务，agent 通过 x402 付 0.05 USDC 后，真实浏览器跑完移动与桌面购买流程并返回截图和失败步骤。
A store can look perfectly fine while its checkout is quietly losing customers. Imagine running an online store. Your product pages load, your images look great, everything seems normal. But on mobile, someone taps "Add to Cart" and nothing happens. No obvious error. No warning. They just leave. That's the problem I wanted to tackle for the @moonpay PayBox hackathon. I built AisleRun, an on-demand checkout testing service that AI agents can pay to use through @PayBox. Instead of asking an AI to guess whether a store works, AisleRun sends an actual browser through the buying journey, testing the same steps a customer would take. Here's what we built and verified: An agent can request a storefront inspection through PayBox's use_service tool. PayBox processes a real $0.05 USDC payment through x402 on Base. The browser inspection begins only after payment settles. AisleRun opens the storefront in Chromium, tests the mobile and desktop buying journeys, and records what happens. The merchant receives a report with the verdict, screenshots, browser recording, trace and the exact step that failed. The interesting part was what happened when we tested it. On our controlled test storefront, AisleRun found a broken mobile Add to Cart interaction and reproduced the failure twice. Mobile: FAIL. Desktop: PASS. We fixed one line of code, commissioned a completely separate paid inspection through PayBox, and ran the tests again. Mobile: PASS. Desktop: PASS. Two separate payments. Two independent browser inspections. An actual failure, followed by a verified repair. We also tested a healthy storefront and bot-blocked behavior. When the browser couldn't reliably inspect a store, AisleRun returned INCONCLUSIVE instead of inventing a defect. The PayBox CLI payment flow is working end to end, with real USDC settlements on Base. The chat-app signing flow and public marketplace listing are still separate integration steps. All storefront examples in the demo are clearly labelled test fixtures, and the recordings come from real browser executions. That's what excites me about agents being able to pay for services: they can commission useful work and bring back something a person can actually inspect. AisleRun is live. Product: https://aislerun.vercel.app Before/after inspections: https://aislerun.vercel.app/demo/compare Documentation: https://aislerun.vercel.app/docs Built for MoonPay Hackathons, Chapter 1. The full three-minute demonstration is attached: