Delegate the purchase — not control of the money
AgentCart turns a natural-language shopping goal into an explainable, budget-bounded PayPal checkout. Give it a goal such as “Set up a home office for $200”: it plans the cart, negotiates when the budget is tight, re-plans around the result, and keeps the buyer in control of approval.
What it does
- Parses the goal and hard budget.
- Builds a cart from in-stock catalog items without duplicates.
- Negotiates a 5% bundle discount on the priciest item when that can unlock a skipped need.
- Substitutes the best-rated same-category item if stock changes.
- Shows a decision trace, then creates a PayPal order and waits for human approval before capture.
- Supports a partial refund after capture.
The budget gate runs before order creation. In the reproducible mock scenario, the negotiated home-office cart totals $199.55 against a $200 budget (withinBudget=true), with 13 trace events and an order result.
Technological Implementation
The app is Express + TypeScript on Node 20. A PayPalGateway interface keeps the same flow across LiveSandboxGateway (PayPal Orders v2 create/approve/capture/refund) and MockGateway (zero credentials, deterministic judge run). The agent can use the default scripted planner or an optional OpenAI/Mistral tool-calling planner over the same typed catalog/cart/order tools. A sandbox hard-lock rejects any PAYPAL_ENV other than sandbox.
Design
The single-page UI puts the goal, live cart, budget boundary, decision trace, and payment state in one timeline. The buyer can inspect the plan and must explicitly approve before money moves; capture and refund are visible state transitions rather than hidden agent actions.
Potential Impact
Delegated shopping is useful only if convenience does not mean blind spending. AgentCart demonstrates a concrete path to fewer manual shopping decisions while preserving a hard budget, human approval, and an auditable explanation of what the agent selected, skipped, negotiated, substituted, and refunded.
Innovation / Idea
The differentiator is the negotiation → re-plan → substitution → refund loop. AgentCart does not stop at recommendations or a cart: it earns room inside the budget, revisits skipped needs, handles a stock change, and keeps every money-moving step behind PayPal approval and the same budget boundary.
Presentation
The demo video shows AgentCart as a working product, not a slide deck: a goal and budget become a traced cart, then a human-approved PayPal payment flow. The public code includes a reproducible mock mode for judges and a documented PayPal Sandbox path for the live integration.
How to run
Clone the repository, use Node 20, run npm ci, then start the demo in its default scripted/mock mode. No API key is required for the reproducible path. Sandbox credentials are needed only for the optional live PayPal Sandbox path; live production payments are intentionally unreachable.
Built with
Express, TypeScript, PayPal Orders v2 REST API, Node.js, OpenAI API, Mistral AI.
What we learned
Agentic commerce is a budgeting and trust problem, not only a browsing problem. The meaningful decisions happen when money runs out: what to drop, what to negotiate, what to substitute, and when to ask the buyer.
What's next
Real merchant catalogs, persistent carts, multi-merchant split payments, and an approval experience tailored for delegated purchases.
Built With
- express.js
- mistral-ai
- node.js
- openai-api
- paypal-orders-v2-rest-api
- typescript