WebAZ Agent Commerce Lab
Point your AI at a real store. Reviewed physical products, honest facts, real payment rails — and a hard rule: no order exists until a human approves it with a Passkey.
✅ Your AI can (no login)
- Search & compare reviewed, active, in-stock physical products
- Read price (USDC display pricing), stock, delivery ETA, return & handling terms
- See decision flags (e.g. new seller, no sales history) and honest unknowns
🚫 It cannot
- Place an order or move funds from this public surface — it is discovery-only
- Accept terms or “buy on your behalf” anywhere without you
- Invent missing facts — unknown compatibility stays unknown
🔑 Where a human approves
- On the authenticated surface an agent may quote → draft → submit a request
- Execution stops there: a human Passkey approval creates the order
- Payment then runs on a real rail you choose — never through the agent
Connect an endpoint
https://webaz.xyz/mcp/shopping-v1https://webaz.xyz/mcpThe shopping endpoint exposes exactly one tool (webaz_search) and cannot create orders or move funds. The full endpoint needs an api_key (a real human registers by invitation with a Passkey; agents cannot self-register).
Client setup
Try these prompts
1 · Phone-stand comparison
Using the WebAZ shopping tool, find phone stands and compare all results in a table: price, stock, delivery ETA to Singapore, return window, handling time, and every decision flag. Separate facts the seller states from facts that are missing, and do not assume compatibility with any phone model unless it is stated. Which would you shortlist for a desk, and why?
2 · Laptop-stand known / unknown test
Search WebAZ for “Metal Laptop Cooling Stand”. Produce two lists: (A) facts the listing actually states, with values; (B) facts a careful buyer would want that are NOT stated (for example exact dimensions, supported laptop size range, maximum load). For list B, say “unknown” — do not fill gaps with typical values for similar products.
3 · Sandbox transaction trace (local, isolated)
Using the local WebAZ sandbox MCP server (stdio, WEBAZ_MODE=sandbox — isolated from the real network), walk one full order lifecycle end to end: create a product, place an order, accept, ship, deliver, confirm. After each step, print the order status and what changed. This is a local simulation; state clearly that nothing touches the live network.
How an agent purchase actually works
Everything left of the key is reversible and money-free. The key step cannot be performed by an agent: it is a WebAuthn Passkey ceremony on a human’s device.
Three modes, clearly separated
| Mode | Credential | What it can touch |
|---|---|---|
| Network read-only | none | Live reviewed catalog via /mcp/shopping-v1 (or the stdio package without a key). Discovery only. |
| Authenticated network | api_key (human-registered) | Quote / draft / submit on the live network. Order creation still requires the human Passkey approval above. |
| Local sandbox | explicit WEBAZ_MODE=sandbox | A local, isolated database on your machine. Full lifecycle simulation; nothing reaches the live network. |
Honest facts, honest unknowns
A real listing from the reviewed catalog — what the surface states, and what it deliberately leaves unknown:
Known (stated by the listing)
- Metal Laptop Cooling Stand · brand Choseal · foldable
- 18.4 USDC · in stock · 7-day returns · 72 h handling
- Delivery ETA: Singapore ~12 days, elsewhere ~18
- Decision flags: new seller (≤90 d) · no sales history yet
Unknown (and reported as unknown)
- Exact dimensions and weight
- Supported laptop size range and maximum load
- Included parts and material thickness
- The surface never fills these gaps with “typical” values — ask the seller before relying on them
Anything beyond discovery — questions to the seller, ordering, payment — happens on the product page in WebAZ itself.