-
-
A unified dashboard showing monthly spending, pending payments, blocked transactions, and recent purchase activity.
-
AI-powered product discovery through Channel3, with prices, retailer details, and spending decisions.
-
PayPal Sandbox checkout with buyer approval and payment capture.
-
Purchases exceeding the automatic transaction limit require human approval before proceeding to PayPal checkout.
-
Persistent purchase records provide visibility into pending and completed transactions processed through the PayPal payment workflow.
-
Alerts users when purchases require approval before checkout.
Inspiration
AI shopping agents are becoming increasingly capable of discovering products and helping users make purchasing decisions. But giving an AI agent the ability to shop also raises an important question: How do users stay in control of what an AI agent spends on their behalf? I wanted to explore this problem by building AgentCart — a trust layer for AI-powered shopping agents. My goal was to combine AI-driven product discovery with spending policies, human approval, and PayPal payments so that convenience would not come at the cost of financial control.
What it does
AgentCart allows users to describe what they want to buy in natural language. Its AI-powered shopping assistant interprets the request and helps discover relevant products through Channel3. Before a purchase can proceed, AgentCart's policy engine evaluates it against the user's spending rules.
Key features include:
- AI-powered shopping: Understands natural-language shopping requests using Groq, with Gemini as a fallback and a local parser as an additional fallback.
- Product discovery: Integrates Channel3 to search for products and retrieve product information.
- Spending controls: Users can configure monthly spending limits and per-transaction limits.
- Purchase decisions: Purchases are classified as allowed, requiring approval, or blocked based on spending policies.
- Human-in-the-loop approvals: Purchases requiring additional authorization go through an approval workflow before checkout.
- PayPal integration: Uses PayPal Sandbox to create orders, obtain buyer approval, and capture payments.
- Transaction history: Records purchase activity and provides visibility into transaction statuses.
- Retailer trust indicators: Recognizes supported retailer domains and flags sellers that require additional verification. AgentCart aims to make AI-assisted shopping more transparent, controlled, and accountable.
How we built it
I built AgentCart as a full-stack JavaScript application.
Frontend: React and Vite provide the shopping interface, spending rules dashboard, purchase requests, and transaction history.
Backend: Node.js and Express handle API requests, product validation, spending policy evaluation, approval workflows, and PayPal integration.
AI: Groq (using GPT-OSS 20B) is the primary model for interpreting shopping requests. Gemini provides a fallback, with a local JavaScript parser as an additional fallback.
Product discovery: Channel3 provides product search and product information.
Payments: PayPal Developer Sandbox is integrated for order creation, buyer approval, and payment capture.
Database: Turso provides cloud-hosted persistence for spending rules and transaction records.
Deployment: The frontend is hosted on Vercel and the backend on Render.
The architecture separates AI intent interpretation from purchase authorization. The AI helps understand what the user wants, while the backend independently evaluates spending policies before allowing a purchase workflow to proceed.
Challenges we ran into
One of the biggest challenges was ensuring that AI-assisted shopping did not mean blindly trusting AI decisions. I had to separate product discovery and intent interpretation from the actual purchase authorization process. The backend needed to revalidate product information, prices, availability, retailer verification and spending rules before proceeding. Integrating PayPal also required understanding the difference between creating an order, obtaining buyer approval, and capturing the payment. A transaction should only be marked completed after PayPal confirms a successful capture. Another challenge was maintaining consistent data across the frontend, backend, and cloud database while deploying the application. Debugging these integrations helped me understand how real full-stack applications coordinate external APIs, payment workflows, and persistent data.
Accomplishments that we're proud of
- Built a working full-stack AI shopping application independently.
- Integrated multiple AI model options for natural-language shopping intent interpretation.
- Implemented a backend spending policy engine with allowed, approval-required, and blocked purchase decisions.
- Built a human approval workflow before checkout for purchases requiring authorization.
- Successfully tested the PayPal Sandbox order, buyer approval, and payment capture flow end to end.
- Implemented persistent spending rules and transaction history using Turso.
- Deployed the frontend and backend so the application can be accessed online. The most meaningful accomplishment is that AgentCart goes beyond recommending products: it demonstrates how an AI shopping experience can incorporate financial rules and human oversight into its payment workflow.
What we learned
Building AgentCart taught me how to integrate AI services into a full-stack application and how to design systems where AI recommendations and financial authorization are separate responsibilities. I gained practical experience with React, Node.js, Express, external APIs, database persistence, PayPal Sandbox payments, cloud deployment, and debugging multi-service applications. Most importantly, I learned that building an AI agent is not only about making it capable of taking actions. It is also about designing boundaries, accountability, and meaningful human control.
What's next for AgentCart — The Trust Layer for AI Shopping Agents
There is room to take AgentCart further. Future improvements could include user authentication and individual spending profiles, more comprehensive retailer verification, additional payment options, stronger security and automated testing, and more advanced agent capabilities. My long-term vision is to make AgentCart a reusable trust and financial-control layer that helps people confidently delegate shopping tasks to AI agents while retaining control over their money.
Built With
- 20b
- channel3
- developer
- express.js
- gemini
- gpt-oss
- groq
- javascript
- node.js
- paypal
- platform
- react
- render
- sqlite
- turso
- vercel
- vite