AI Agents Can Find Products. The Hard Part Starts After "Buy."

WebAZ, an open-source protocol, aims to make the post-purchase process in AI agent commerce explicit and inspectable. It now runs two settlement rails, including a USDC lock on Base mainnet, without taking custody of funds. The protocol also adds operational features for sellers and supports both human and agent interfaces.

AI agents can already search catalogs, compare options, and recommend a product. That is useful, but it is not commerce. Commerce begins when the recommendation becomes a commitment: Most agent demos stop before these questions. WebAZ is an open-source protocol trying to make the messy part after "buy" explicit, inspectable, and usable by both humans and AI agents. WebAZ now runs two real settlement rails. On neither rail does WebAZ take custody of the principal. The buyer pays the seller directly, off-platform. WebAZ records the payment-instruction snapshot, risk acknowledgement, order state, and evidence, but it does not verify the payment method, guarantee the payment, or issue the seller's refund. This path is intentionally simple for sellers who already have a preferred way to receive money. Real USDC can be locked in an immutable contract on Base mainnet. WebAZ does not hold a key that can redirect those funds to an arbitrary address. The contract constrains exits to the original buyer, the voucher-bound seller, or the fee sink, and enforces a per-order amount cap on-chain. An important boundary: the contract has not had a third-party security audit. The current compensating controls are a deliberately small immutable surface, hard transaction caps, a deposits-only pause, and enumerated fund exits. This is a live rail, not a claim of zero risk. Payment is only one part of a purchase. Recent WebAZ releases also added operational pieces that independent sellers need: These are not flashy agent demos. They are the state transitions that determine whether a real order can be understood later. Humans can use the WebAZ PWA. Agents can connect through the live Remote MCP endpoint. Public product discovery is available anonymously. Authenticated writes can use OAuth 2.1 authorization code flow with PKCE, while selected higher-risk actions retain a Passkey approval boundary. A fixed shopping MCP surface also gives clients a narrow product-discovery interface instead of exposing every available tool. The goal is not to remove the human from commerce. It is to make the handoff between human intent and agent execution visible. The newest merged work includes a dormant, read-only dual-RPC shadow observer for a future escrow path. It cannot write to the chain, has no signer, changes no money or order semantics, and is not connected to production routing. That distinction matters. Infrastructure can be tested against live conditions without quietly becoming a live payment path. Agent-native commerce will need more than a shopping chatbot and a payment button. It needs shared state, explicit authority, bounded money movement, fulfillment evidence, and recovery paths that another human or agent can inspect. WebAZ is building that protocol in the open, one complete state transition at a time. Which part of the post-purchase loop is hardest in your own agent or commerce workflow: approval, settlement, fulfillment evidence, or returns?