Standing Orders

Standing Orders is an AI agent that watches product prices for you, buys genuine deals through PayPal, and refuses fake "sales" — with spending caps and a kill switch it can't override.

Comment

Inspiration

Online "sales" are often a price quietly raised, then discounted — and nobody has time to watch prices for a real deal. We wanted an agent you could trust with a budget: one that judges deals, transacts on its own, and still has the brakes a human would want.

What it does

You describe what you want in plain English and a max price — "a 27-inch 4K monitor under $350." Claude turns it into a structured spec. The agent then searches live products via Channel3, scores each for fit, and decides one of four actions:

  • Fill — a genuine deal, under your cap, from an approved merchant → buy through PayPal.
  • Hold — it fits, but a better price is likely soon → keep watching.
  • Skip — the "discount" is measured from an inflated price → refuse the fake sale.
  • Ask — it fits but needs your sign-off → send a one-tap PayPal approval.

Spending caps, a merchant allowlist, and a kill switch are enforced in plain code the model cannot override. Purchases complete through the PayPal sandbox with real order creation and capture.

How we built it

  • Next.js + PostgreSQL for the app and data.
  • Claude (via the Vercel AI SDK) does three jobs — parse intent, score candidates, decide — each call audit-logged.
  • Channel3 supplies live product data across millions of items.
  • PayPal Orders API creates and captures real sandbox payments; a one-tap approval flow completes each purchase.
  • Render runs it as a blueprint: web service, Postgres, and a cron job that hits /api/tick every 15 minutes to advance active orders — that's what makes an order "standing."

Challenges we ran into

The hardest design decision was keeping judgment separate from enforcement: the model proposes an action, but plain code enforces the spending caps and kill switch and is the only thing allowed to touch PayPal. Fake-sale detection also needs real price history, so the agent records prices on every check and reasons over the trend rather than trusting a single advertised discount.

Accomplishments that we're proud of

  • An agent that visibly refuses a fake sale — it catches a discount measured from a recently inflated price and explains why, in plain language.
  • A genuine end-to-end purchase: the agent decides to buy, the user taps approve, and PayPal captures the payment.
  • A clean separation of concerns — the model's reasoning is always visible, but it can never spend money the guardrails don't allow.

What we learned

That the interesting part of agentic commerce isn't getting an agent to buy — it's getting it to not buy for the right reasons. Most of the value came from the hold and skip paths and from making every decision auditable, so a human can always see what the agent thought and why.

What's next for Standing Orders

Agent-ready merchant checkout beyond our sandbox store, real notifications for approvals, and live fake-sale detection once an item has accrued a couple of weeks of recorded price history.

Built With

Share this project:

Updates