SnapFlip — sold before you buy it
An order book for secondhand goods where AI agents bid on your thrift find while it's still on the rack.
Live: https://go.snapflip.workers.dev — landing page with real-time demand; /buy to set a standing order, /mcp?token=… as a Claude connector, /studio for the 3D inspection experience. Text the collector desk: +1 650 315 6536 — Scout qualifies you, cites the live book, and hands off a prefilled order. Just browsing? Text it the thing you lost as a kid: it prices the buy-back off real market data (/api/market), appraises your whole childhood box (/box), and sets a price watch that pings you on Telegram when the tape dips (/watch).
The problem
Resellers carry all the risk. They buy inventory on a hunch, spend time listing it, and hope it sells. Tools like eBay's AI lister, Vendoo, and Poshmark save typing. None of them tells you, at the rack, whether a real buyer will pay for the item.
Collectors have the opposite problem. They know exactly what they want and what they'd pay, but they have to keep refreshing marketplaces to find it.
How it works
- Buyers set standing orders with an agent. "Pokémon Yellow, cartridge only, authentic, good label, up to $45." They can do this through Claude (MCP), a Brainbase-hosted concierge, the web, or by texting Scout — our collector-desk agent on OpenClaw (
scout/): +1 650 315 6536. The buyer grants a spending limit that only works at SnapFlip, is capped, and expires.
- A reseller snaps an item at the thrift store (Telegram). Claude identifies the exact SKU and grades its condition from the photo — and shows its work: confidence, the candidates it weighed, and evidence lines for the grade, marked on the photo itself.
- A flash auction starts among every buyer agent whose order matches. It's a 60-second ascending-clock auction. The price ticks up, and each agent drops out once the price passes what its buyer would pay for this specific item and grade. The seller watches the bids climb live on their phone.
- The seller decides with certainty. "Cleared at $38. You net $34.20 after fees. Rack price $6." They buy the item, confirm with a photo, and the buyer's card is charged.
- The item ships, and funds release to the seller on delivery (Stripe Connect).
If the price doesn't clear the seller's reserve, they walk away and have lost nothing.
Why it's different
- Pre-sold sourcing. The reseller's decision to buy happens after price discovery, not before.
- Demand first. Listings don't need traffic. The buyers are already waiting with money committed.
- Agents with real authority. Buyer agents commit money on their own, within limits their human set. That's autonomy with guardrails, not a chatbot.
- Auctions that feel native to agents. A clock auction means each agent just needs to know its buyer's true value. Sniping and bid-shading don't help, so agents stay simple and honest, and it's a live show for the seller.
SnapFlip Studio
/studio is a second product built on the same backend: an AI-native 3D inspection-and-auction experience. A seller's photos become a spatially inspectable object; a draggable raking lamp reveals photographed condition; and simulated buyer agents make decisions that cite the evidence — running as one continuous Three.js world across three chapters (the Rack, the Inspection Atelier, the Auction Chamber). WebGPU where available, WebGL2 otherwise.
Strategy, gates and acceptance tests: docs/studio/STRATEGY.md. What's actually built and verified: docs/studio/STATUS.md.
Business model
A 10% seller fee on the clearing price (SELLER_FEE_BPS, snapshotted per sale into the platform_fees ledger so rate changes never rewrite history), against typically 13–15% all-in on eBay and 20% on Poshmark. Sellers accept it because the item is already sold — the fee is quoted live on every auction page and broken out in the seller's payout message. Later: a buyer subscription for priority matching, and data products built on real-time demand.
Wedge
Physical collectibles with an exact identity — retro video games first. Every item has a canonical SKU, collectors already keep wishlists, carts turn up constantly at thrift stores, and counterfeit risk is manageable (the model flags reproductions, code disqualifies them). The engine is category-agnostic: the catalog is a table, so vinyl records (exact pressings), LEGO (set numbers) and TCG (card + set) are a data migration, not a rebuild. Sneakers later, once authentication exists.
Quickstart
npm install
cp .dev.vars.example .dev.vars # fill in your own keys
npm run db:local # local D1: schema + seed
npm run dev # wrangler dev (builds the studio bundle first)
- Telegram: make your own test bot with BotFather, expose dev with
cloudflared tunnel --url http://localhost:8787, then npm run tg:webhook -- <tunnel-url>.
- Verify before pushing:
npm run check (studio typecheck + tsc + wrangler dry-run). Also npm test, npm run lint.
- Deploys:
npm run deploy — but check npx wrangler whoami first; production lives on a specific Cloudflare account (see AGENTS.md).
Stack
| Provider | Used for |
|---|
| Anthropic | Claude vision (SKU identification and condition grading), match verification, parsing plain-English order rules and valuing each item against them (Featherless as the OpenAI-compatible fallback provider) |
| Cloudflare | Workers, Agents SDK (one Durable Object per buyer agent), a Durable Object per auction (alarms for the clock, hibernatable WebSockets for live bids), remote MCP server, D1, R2, Queues |
| Stripe | Buyer spending limits (saved card charged off-session, cap enforced in code; Shared Payment Tokens slot in later), manual capture, Connect Express seller payouts via separate charges and transfers, webhooks |
| Brainbase | Hosted buyer concierge worker (chat deployment) that creates standing orders through our API |
| Three.js | Studio world — WebGPURenderer (WebGPU/WebGL2 backends), Gaussian splats, post-processing chain |
Docs
| Doc | What |
|---|
docs/ARCHITECTURE.md | System design: components, endpoints, payments flow, schema, trust & safety |
docs/studio/STRATEGY.md | Studio plan: positioning, chapters, gates, acceptance tests |
docs/studio/STATUS.md | Studio running status — what's built and verified |
docs/studio/USER_TEST_PLAN.md | Studio user-test script, clip plan, analytics |
scout/README.md | Scout, the collector-desk agent: modes, safety model, deploy |
docs/history/ | Hackathon-day artifacts: build plan, demo script, Devpost copy |
AGENTS.md has the agent-facing notes (deploy account, studio commands, deep links). docs/internal/ is machine-local ops and never committed.
Originally built for the Startup Speedrun Hackathon (Brainbase × Anthropic × Cloudflare × Stripe), Agentic Payments track. MIT licensed.
