Inspiration
AI shopping agents are racing to "buy faster", but most people don't trust an AI with their money. They want hard spending limits, a way to cancel, prices that weren't made up, and no sponsored bias. We wanted an agent that earns that trust, with PayPal as the safety layer.
What it does
Dalil finds the best deal for a product from a name, a link or a photo, then:
- Signed prices: every offer is signed by the server with the price it actually saw, so the AI can't invent or change one.
- Store trust and fake-deal checks, plus a "buy now or wait?" verdict from price history.
- Spending mandate: your limit is enforced in server code when money moves. Even if the AI proposes it, an over-limit payment is refused.
- Agent-to-agent negotiation: Dalil haggles with the store's own seller agent and returns a lower, signed price.
- PayPal hold → capture: funds are authorized and held, you get a 10-minute cancel window, then capture, and a trust receipt.
- Price protection: for 14 days, if the price drops, the difference comes back automatically as a partial PayPal refund.
- Price watch: target price, background checks, alert or auto-buy within your mandate (Vault).
- Multi-store cart: one PayPal order and one approval for several stores, with a per-store refund.
- Family approval: a parent or partner approves and pays from their own PayPal, within a limit you set.
- Group buy: split a purchase with friends; everyone's share is held until all have paid.
- Guided demo: one button runs every feature on the live sandbox with the real AI.
How we built it
- Next.js 15 on Vercel.
- Meta Muse Spark (muse-spark-1.3-contributor) runs a tool-calling agent loop over the Channel3 product catalog, web search, store pages and a real headless Chromium.
- PayPal Orders v2: AUTHORIZE for single purchases, multi-purchase-unit CAPTURE orders for carts, and shipment tracking.
- PayPal Payments v2: capture, void, and full and partial refunds.
- PayPal Vault v3 for auto-buy. All money-moving calls use idempotency keys.
- Upstash Redis stores watches, protections, family and group requests. GitHub Actions checks prices every 10 minutes. AG Grid shows the purchases and savings dashboard.
Challenges we ran into
- Multi-store cart: real PayPal rejects AUTHORIZE for orders with several purchase units, so we redesigned the cart as one CAPTURE order with a capture and a refund path per store.
- Store trust from catalog links: catalog links go through a redirect domain, so we resolve each store's real domain from Channel3's own data on the server, never from the model.
- Browsers block popups: PayPal can only open from a real click, so the guided demo pauses for the judge at payment steps.
Accomplishments that we're proud of
- Every guarantee is enforced by server code, not by the prompt.
- The full flow works end to end on the live site with the real model and the real PayPal sandbox: the 12-step guided demo passes, and our API and attack suite passes 113/113 on production.
What we learned
Trust in agentic commerce comes from verifiable steps. Signed prices, server-side mandates, holds before capture and automatic refunds turn "trust the AI" into "check the AI".
What's next for Dalil
- Real merchant seller agents.
- More markets and currencies.
- Live (non-sandbox) PayPal after a security review.
Built With
- ag-grid
- ai-agents
- channel3
- chromium
- github-actions
- meta-muse-spark
- next.js
- paypal
- paypal-orders-v2
- paypal-payments
- paypal-vault
- react
- tailwindcss
- typescript
- upstash-redis
- vercel