Inspiration

Everyone has a signature niche: the pattern, color or obsession they'd buy anything in. Mine is houndstooth. But a niche shop has always meant sourcing products, building a site and handling payments, which is too much for most people. I wanted to see if an AI agent could do all of that from a few words and if PayPal could carry the money end to end: the founder paying to launch, shoppers paying merchants and founders getting paid.

What it does

Curatique turns any niche into a real, shoppable store, curated by AI and paid with PayPal.

For shoppers

  • Branded niche storefronts (flagship: Houndstooth etc.) of live products from Channel3's catalog
  • PayPal-first ranking: merchants that accept PayPal rank first and carry a Pay with PayPal badge. Others still appear, just lower.
  • A Claude shopping concierge with real tools (search, product details, similar items). Every step it takes is visible under Agent steps.
  • Product pages request offers from PayPal merchants only, so the Buy button goes where you can check out with PayPal.

For founders

  • Type a niche. In about 35 seconds Claude designs the brand and theme, plans dozens of Channel3 searches and keeps only on-niche items, each with a curator note.
  • Pay the $27 launch fee with PayPal Checkout (Orders v2) and the store goes LIVE.
  • Earn affiliate commissions, paid out with PayPal Payouts. Every buy link is tagged to its store for attribution.

How we built it

  • Next.js 15 + TypeScript + Tailwind, hosted on Render with a persistent disk
  • Claude, with two models for two jobs:
    • Haiku 5.5 for the fast, frequent calls: the concierge's tool-use loop and PayPal merchant checks, where it uses server-side web search and web fetch to read a retailer's payment FAQ and return a verdict with evidence
    • Sonnet 5.5 for judgment: brand design, search planning, and curation, using structured outputs
  • Channel3 API: /search with pagination and audience filters, /products/{id} with website_ids to force a PayPal-merchant offer, and /similar for recommendations
  • PayPal REST (sandbox): Orders v2 with the JS SDK buttons and server-side capture, Payouts v1 batches with status polling, and OAuth client credentials

Ranking blends Channel3 relevance with a PayPal boost:

$$\text{score} = \text{relevance} + b_{\text{PayPal}}, \quad b \in {0.6 \text{ accepts},\ 0.15 \text{ unknown},\ 0 \text{ no}}$$

Challenges we ran into

  • Channel3 doesn't say who accepts PayPal. So I built a cached merchant table, seeded it with known retailers and had a Claude agent with web tools research unknown domains live and cache the verdict.
  • Cheap models drift off-niche. Haiku let about 1 in 3 off-niche items (herringbone in a houndstooth store!) into the catalog. Moving curation to Sonnet made every pick on-niche, so I split the work: Sonnet curates, Haiku chats.
  • Kids' items slipping in. Searching "holiday dresses" returned children's clothing. I wrote audience rules: dresses, skirts and heels default to adult women, and suits and boxers to adult men, unless kids, girls or boys are named. The rules are enforced in Channel3 filters, in a post-filter and in the prompts.
  • Bad data. A $1 listing for a designer item led me to drop offers priced below 20% of the item's median.
  • Honest demo numbers. No real sales exist yet, so earnings are simulated and clearly labeled. But only real affiliate items "sell", at their real Channel3 rate, after Channel3's 50% share:

$$\text{commission} = \text{order} \times \text{rate}_{\text{Channel3}} \times 0.5$$

Accomplishments that we're proud of

  • Everything in the demo is real and live: Channel3 calls, Claude agents, PayPal Checkout and Payouts, all at https://curatique.sha-i.com
  • A complete store from a single niche in about 35 seconds
  • PayPal is part of every flow: the launch fee, shopping and payouts
  • Agent transparency: shoppers can see every tool call the concierge makes

What we learned

  • Match the model to the job. A small model is fine for conversation; curation needs judgment.
  • Agents are most trustworthy when they cite evidence, as the merchant checks do and show their steps, as the concierge does.
  • Honest, clearly labeled demo data beats impressive-looking made-up numbers.

What makes Curatique different?

AI store builders are a crowded category, but nearly all of them generate dropshipping stores: you pick products, manage suppliers and fulfillment and carry the inventory risk. Curatique works differently:

  • No inventory, no fulfillment. Curatique curates real products from thousands of retailers through Channel3. The retailer sells and ships; the founder just curates.
  • Affiliate, not dropshipping. Founders earn commission on purchases made through their store's links, not margin on goods they have to stock, so there's nothing to buy up front.
  • PayPal at the center. Merchants that accept PayPal rank first and carry a badge, the launch fee is paid with PayPal Checkout and founder earnings go out through PayPal Payouts.

The closest comparison is creator storefronts like LTK and ShopMy, but those are hand-curated by influencers who already have an audience. Curatique is an AI that builds your niche affiliate store from across the whole internet: type a few words and in about 35 seconds you have a branded, shoppable store with an AI concierge, no following required.

What's next for Curatique

  • Real earnings from Channel3's reporting API feeding PayPal Payouts automatically
  • A custom subdomain for every store, plus founder editing (reorder, pin, remove items)
  • More PayPal: Venmo and Pay Later messaging on product pages, and founder payouts on a schedule
  • Taste-learning, so the concierge remembers what each shopper loves within the niche

Built With

  • ai-agents
  • anthropic-api
  • channel3
  • claude
  • claude-haiku
  • claude-sonnet
  • json
  • next.js
  • node.js
  • paypal
  • paypal-checkout
  • paypal-orders-api
  • paypal-payouts
  • paypal-sandbox
  • react
  • render
  • rest-api
  • structured-outputs
  • tailwindcss
  • tool-use
  • typescript
  • web-search
Share this project:

Updates

Submission history