Searching the stockroom alone or asking at the counter.
Picture a store where the customer has to walk into the stockroom and find the product alone, eyes almost shut, feeling shelf by shelf. It works, but it is slow, and moving one aisle is enough to get them lost. That is how an AI assistant shops today: it takes a screenshot, reads the code of the whole page, guesses where the button is and clicks.
The other way is the counter: the store has a menu of what the clerk can fetch, and the customer just asks. WebMCP is that counter for agents. The store itself declares the functions an agent may use, with a name and a description, and the agent calls the function instead of imitating a person.
We call these functions sockets: a standard fitting, the same for any agent, that the store switches on or off.
Screenshot, read, click: four ways to go wrong.
It costs at every step. Each action forces the agent to look at the whole screen again. In a ten-step purchase that multiplies, and whoever pays the AI bill feels it.
It breaks when the layout changes. The team moves a button and the automation stops. Fixing it becomes a permanent tax.
Hidden text can give orders. Third-party content on the page (a comment, an ad) can carry instructions telling the agent to act against the store or the customer.
Being able to click is not permission. The agent does whatever the screen allows, even without authorization for the business rule behind it. Whoever confuses the two opens a gap.
The store declares what the agent may ask for.
There are two ways to open a socket. In the simplest, the store marks a form that already exists, such as search or postcode, with a name and a description, and the agent starts using it as a function. In the other, the site registers a function in the page’s own code. Neither requires rebuilding the site.
The standard also brings safety hints: marking content as "untrusted" and a function as "read-only". They help, but they are a patch: the layer that lets the agent act arrived before the layer that governs what it may do.
Clicking is not permission. The store decides.
Every socket that touches money or personal data goes through three locks: the store’s rules on the server (value cap, stock, fraud check), a human confirmation gate for anything irreversible, and a log of each step.
The rule we take from real 2026 incidents: everything the agent reads is suspect, and the gate that matters is the exit. Trying to clean everything that comes in does not work; it is safer to control what the agent can close, pay for or send out.
Sheet for each socket
Your store receives agents. Your robots visit other stores.
In your store (receiving)
- Open read-only first: catalog, price, stock, shipping, order status.
- Then what reserves: cart, with a short expiry.
- Last, what charges: place the order, always with a confirmation gate and a cap.
- Make yourself findable: a public catalog of your sockets helps agents find you; being found is not being trusted.
In your automations (visiting)
- Use the socket where it exists: and fall back to reading the screen where it does not.
- Isolate the method: a layer in the middle turns the switch into an upgrade, not a rewrite.
- Measure cost per action: screen versus direct call, on your own workload.
- Treat the page as hostile: read, never obey.
A function the store declares for agents to use directly, with a name and a description, instead of the agent imitating clicks.
The step where a person approves anything irreversible, such as paying. The agent proposes; the person confirms.
Text on the page, invisible or disguised, that tries to give the agent orders. That is why it reads but does not obey.
Where this may be wrong.
The standard is still on trial
WebMCP is in public trial in Chrome and in preview at Cloudflare. Names, formats and rules may change before it becomes official. That is why we recommend the layer in the middle, which absorbs the change.
We did not measure the gain
Calling a function costs far less than reading the whole screen, but we have no number we measured for Brazilian stores. The specimen counts actions, not cost.
Few agents use it today
Opening the socket does not bring traffic by itself. It pays off when the assistants your customers use start looking for it.
Security is behind
The standard’s safety hints are stopgaps. Real protection is still the rule on the server, the human gate and the log.
← Agent-ready e-commerce · stickybit.com.br
- InfoQ, "WebMCP Standard Proposal for Agentic Web Actuation Now Available in Chrome (Origin Trials)" (13 Jun 2026)
- InfoQ, "CloudFlare Previews Automatic WebMCP Support for Web Pages" (10 Aug 2026)
- InfoQ, the ARD spec for agent tool discovery (Jul 2026)
- Bloomberg, Amazon blocking Muse (21 Sep 2026)
- The rule "every channel the agent reads is hostile; the gate is the exit": our reading of three July 2026 incidents (internal notes)