The Problem

AI can already find good prices, but only about 1 in 10 people will let an agent finish a purchase. The fear is reasonable: studies show agents buying from fake stores, trusting fake reviews, and getting hijacked by hidden text in listings. Payment rails for agents already exist. What's missing is a reason to trust the agent with the money.

Our Value

Hunter reimagines shopping: you hand a purchase to an AI without handing it your wallet. You say what you want and set a limit once. Hunter searches real marketplaces, asks only the questions that matter, and buys only after you confirm.

Speed. Marketplaces are built for people, with logins, carts and checkout forms. So we rebuilt three for agents: we scraped 2,000+ real product listings across Amazon, eBay and Depop with Apify and serve them from our own storefronts. Each one sells through x402, so an agent pays over plain HTTP with no account and no checkout form. The first listing arrives in under a second, and a full search with safety screening finishes in 5–12 s.
Security. The language model never touches the money. A deterministic payment Gate checks the store, the seller's wallet, the price and your limit before anything is signed. A prompt injection can change what the model says. It can't change what the Gate approves.

We attack our own Gate. A Factory Droid coordinator runs red-team rounds. Bitdeer models write new scams from real listings, the Gate retrains on whatever slips through, and a new model ships only if scores computed in code improve. Fixed rules catch lookalike wallets, price switches at checkout, and unknown sellers over a $100 cap. The red-teamed model catches lookalike stores, too-cheap deals, prompt injection, hidden text and fake reviews.

How it Works

Understand: the Scout, running on Bitdeer AI Model Studio (GLM-5.3-Flash), turns "a warm fleece under $50" into a query and a limit, and streams its reasoning live.
Search: our Apify Standby Actor crawls every storefront in parallel and streams listings as they're found. It works on any x402 store that publishes its catalog, and it pulls out hidden text so the Gate can see what a human can't.
Screen: the Gate scores every listing while the search streams. Layer 1 is fixed rules; Layer 2 is a risk model that can only block, never approve.
Pay: after you confirm twice, the Gate re-checks the store's real 402 Payment Required response. The agent then signs a USDC transfer for exactly that amount, and Coinbase's x402 facilitator settles it on Base.

Results

On the live stores, the Gate blocked 105 of 105 scam-store listings and 0 of 304 real ones.
On 444 test listings it never trained on: 90.7% accuracy, with only 1.8% of honest listings wrongly blocked.
Red-teaming more than doubled the scams the Gate catches: fake-review scams went from 35% to 85% blocked and prompt-injection scams from 45% to 85%, with no rise in honest listings blocked.