I'm not sure this was ready to show HN unless I am missing something.
The part I care most about: the marketplace never touches the money. A buyer POSTs, gets an HTTP 402 with the payment terms, signs an EIP-3009 authorization, and re-sends it in a header (x402). The transfer is wallet-to-wallet USDC on Base, gasless for the buyer, and our wallet is simply not in the settlement path. You can check that on-chain. Fees are 0% right now, so the 402 carries a single leg to the supplier; when I turn fees on it becomes two legs and the Agent Card advertises the rate so buyers can adapt.
Why non-custodial: I'm a solo operator in Japan, where holding other people's stablecoins is a licensed activity. Removing custody from the architecture was cheaper than acquiring a license, and it turned out to be the better design anyway.
The other piece is quality. Neither x402 Bazaar nor Google's A2A verifies supplier output. They standardize discovery and envelopes and leave quality to reputation. So there's a verification layer here: outputs are scored against JSON Schema, grounded against the source text by label proximity (so a supplier can't pass the management fee off as the rent), and checked against declarative constraints. That score, weighted by real payments and distinct payers, is the reputation. It only works for decidable field types like numbers, enums and dates, which is exactly why I scoped supply to extraction rather than free-text generation.
Registration is permissionless: POST an agent card with an endpoint and a payout address and you're live. The catalog is small and still mostly my own reference suppliers, which is the honest cold-start state.
Feedback welcome, especially on the verification layer and on whether the two-leg payment extension is a dealbreaker for standard x402 clients.