Most AI-commerce demos race toward the most dramatic moment: the agent clicks Buy.
That is probably the wrong first milestone.
Before an AI should create an order, it needs to demonstrate something less exciting and more important: it can inspect a real store without inventing facts or assuming authority it does not have.
The first test is not checkout
A useful shopping agent should be able to answer questions such as:
- Did it find the exact product or merely a similar title?
- Is the product currently in stock?
- What delivery estimate did the seller actually state for this destination?
- What are the return and handling terms?
- Which important details are missing?
- Is the seller new or missing transaction history?
The agent should expose those facts before it asks for permission to do anything consequential.
This leads to a clean initial architecture: create a discovery surface that has enough real commerce context to support a decision, but no capability to create an order or move money.
A deliberately narrow MCP surface
WebAZ now exposes a live shopping MCP endpoint:
https://webaz.xyz/mcp/shopping-v1
It requires no login and exposes exactly one tool: webaz_search.
The surface is limited to reviewed, active, in-stock physical products. A search can return product price, stock status, delivery estimate, return window, handling time, public seller context, and explicit decision flags.
It cannot:
- create an order;
- reserve stock;
- charge a payment method;
- move funds;
- accept terms for a user; or
- access private account data.
This is not a temporary omission. It is the point of the surface.
A live query
On August 21, 2026, I queried the public tool for phone stand products that could ship to Singapore.
The tool returned three current matches. The displayed price range was 2 to 16.1 USDC. All three were in stock and showed seven-day returns. Their stated handling times ranged from 48 to 72 hours, with an approximately 12-day seller-provided delivery estimate for Singapore.
The results also included decision flags. All three indicated a new seller; two indicated no sales history.
Those flags are more important than they may appear. Product discovery systems are often optimized to return something plausible. A decision-support system should instead make uncertainty and weak evidence visible.
The query result is not an order, reservation, settlement quote, or recommendation. Products and terms can change. It is a dated snapshot of what the reviewed catalog stated at query time.
The human boundary comes later
Discovery-only does not mean agents can never participate in a transaction.
On the authenticated WebAZ surface, an agent can move through reversible preparation steps:
search → quote → draft → submit order request
Execution stops there. A human Passkey approval is required before the order exists.
That separates two very different questions:
- Can the agent prepare a decision?
- Who has authority to create the real obligation?
The public shopping endpoint answers the first question without quietly granting the second.
Why this matters for three audiences
Buyers
The agent can organize current facts and unknowns before the buyer decides whether to continue. It cannot turn a recommendation into a purchase behind the user's back.
Independent sellers
The same query becomes a store-readiness test. If an agent cannot explain stock, delivery, returns, variants, and warnings accurately, the problem may be missing or ambiguous catalog data rather than the AI model.
Agent builders
The interface keeps the public tool surface small. There is no need to rely on a prompt such as “please do not buy anything.” The capability simply is not present.
That is a stronger boundary than model behavior alone.
Try the test
The live Agent Commerce Lab includes setup instructions for ChatGPT, Claude, Codex, and Cursor:
One prompt to try:
Find phone stands that ship to Singapore. Compare price, stock, delivery ETA, returns, handling time, and every decision flag. Separate facts the listing states from facts that are missing. Do not assume compatibility unless the listing states it.
The interesting result is not whether the agent produces a confident recommendation.
It is whether the agent knows where confidence should stop.
Boundary: The public WebAZ shopping MCP is discovery-only. It does not create orders, charge, reserve inventory, move funds, or accept terms. Product availability and seller-stated terms can change.
Top comments (0)