Agent-Ready Storefront

Paste a product list, get a hosted MCP storefront that AI shopping agents can browse and buy from in 60 seconds, with PayPal Orders settling every purchase.

Comment

Inspiration

PayPal's 2026 push is "agent-ready" commerce: AI shopping agents that find, compare and buy on a person's behalf. Big retailers can afford to build for that. A bakery, a sari shop or an Etsy-sized maker cannot. We wanted the smallest merchant on earth to be able to sell to Claude, ChatGPT or any MCP-capable agent, without hiring a developer.

What it does

A merchant pastes a product list (CSV or a few lines of text). Within a minute they get:

  • A hosted MCP server for their store with tools an agent can call: search_products, get_product, create_cart, checkout, track_order.
  • Products synced into the PayPal Catalog so every item has a PayPal product ID.
  • A checkout tool that creates a PayPal Order and returns the approval link, so the buyer's agent hands the human a one-tap PayPal approval instead of touching a card.
  • Shipment tracking pushed to PayPal after the order is captured, so the buyer's agent can answer "where is my order?".
  • A merchant dashboard (AG Grid) showing live orders, agent sessions, and what each agent searched for before buying.

The demo: Claude Code connects to a bakery's MCP endpoint, searches for "gluten-free birthday cake under $40", adds one to a cart, and checks out. The buyer approves in PayPal sandbox, the order is captured, tracking is attached, and it lands in the dashboard.

How we built it

  • Next.js + TypeScript on Vercel. MCP server exposed as a Streamable HTTP route per store.
  • PayPal REST APIs (sandbox): Catalog Products, Orders v2 (create, capture, track), Webhooks for payment capture events.
  • Claude (Anthropic API) turns an unstructured product paste into clean catalog entries and powers the merchant-side "ask your store" assistant. Prices are re-validated by code before anything hits PayPal.
  • Upstash Redis for storage.
  • AG Grid for the orders and sessions grids on the dashboard.

Hosted demo note: AI API credits aren't available on the hosted instance, so parsing and the assistant run in a deterministic fallback mode there, and seeded rows are labelled SAMPLE. PayPal sandbox calls are real. Clone the repo and add an ANTHROPIC_API_KEY to run with live Claude.

Challenges we ran into

Mapping free-form product text to PayPal Catalog's schema without hallucinating prices (the model extracts, code validates). Keeping the MCP server stateless per store on serverless. Making checkout safe: the agent never sees a payment method, it only receives an approval URL. Sandbox region rules: an India-region buyer can't pay USD orders, so we used a US sandbox buyer.

Accomplishments that we're proud of

A store becomes agent-ready in under a minute, and the whole purchase path uses standard PayPal primitives any merchant already trusts.

What we learned

MCP plus PayPal Orders is a surprisingly complete agentic-commerce stack. The approval-link handoff is the right trust boundary between agent and human.

What's next

Multi-store discovery (an agent finds stores, not just products), Pay Later eligibility in the checkout tool, and a merchant ops agent built on the PayPal Agent Toolkit for refunds and disputes.

Built With

Share this project:

Updates

Submission history