Inspiration
Most payment problems I hear about aren't clever hacks. Someone pays the wrong number, a "PayPal support" email asks for a gift card, or money leaves your account and sits in pending for two days with no explanation. I work in security, so I wanted a payment app that checks things before the money moves, not after.
Then I got curious about the other side. If an agent can check a payment, can it also do the boring parts of buying and selling? Writing the listing, guessing a price, going back and forth with a buyer. That turned into the second half of PayMinute.
What it does
- You log in with PayPal and pay anyone from one box, a phone number or a PayPal email. AI Guard checks the number, scores the risk (scam phrases, look-alike domains, amounts that are unusual for you, duplicates) and warns or blocks before anything is sent.
- Sell: describe the item in one line. The agent pulls out the features and condition and suggests a price based on similar listings. You can raise it, and your lowest price stays private.
- Buy: say what you want and your budget. Your agent compares every listing feature by feature, then bargains with the seller's agent live. The seller's agent is stubborn on purpose. When they agree, you get a checkout where the seller and price are locked on the server and expire in 10 minutes.
- Payment agents pay a contractor when a job is marked done, or on a schedule, inside limits you set. Anything bigger waits for you.
- AI Guard chat understands plain requests like "split $90 dinner with Priya" or "why hasn't my payment arrived", and uses PayPal's Agent Toolkit and MCP server for the PayPal side.
- Pending payments are explained in plain words, and your PayPal profile, balance and payment requests are shown in the app. ## How we built it The backend is FastAPI with SQLite, and the frontend is React, TypeScript, Vite and Tailwind. On the PayPal side I used Log in with PayPal, Payouts, Orders v2 (authorize, then capture), Invoicing, Transaction Search, Balances and webhooks, all in the sandbox.
The AI runs locally. A scikit-learn Isolation Forest learns each user's own payment history to flag unusual payments, a small TF-IDF and logistic regression model reads chat requests, and the buyer and seller agents are rule-based. I chose not to depend on an external LLM, so every decision that touches money can be explained. The PayPal Agent Toolkit and PayPal's sandbox MCP server handle invoices, orders and transactions from chat. The Studio dashboards use AG Studio from AG Grid.
Challenges we ran into
- Making the bargaining feel real. My first buyer agent paid full price whenever it could afford it, and the seller gave in too quickly. It took a few rounds of tuning to get a seller that holds out and a buyer that uses missing features and the fair price as leverage.
- Checkout security. At one point the seller's private lowest price was leaking into the buyer's results. Now the payee and price live only on the server, a quote belongs to one buyer, and a listing is reserved so it can't be sold twice.
- Sandbox captures stuck on "pending review" even though the money had arrived. I now cross-check with Transaction Search and settle them.
- Keeping accounts apart. Each user only sees their own payments, and the person on the other side sees a mirrored view. ## Accomplishments that we're proud of Watching two agents bargain and close a deal that reads like a real chat. And every payment has a reason you can actually read, instead of just a score. ## What we learned The PayPal sandbox behaves a lot like production, including the annoying parts like payment review holds. And for anything that moves money, simple rules you can explain plus a small model worked better for me than a black box. ## What's next for PayMinute Buyer protection on agent deals, letting the seller agent learn from past sales, and trying it with a few freelancers I know.
Built With
- ag-grid
- fastapi
- log-in-with-paypal
- paypal
- paypal-agent-toolkit
- paypal-mcp-server
- paypal-orders
- paypal-payouts
- python
- react
- scikit-learn
- sqlite
- tailwindcss
- typescript
- vite