agenteconomy.cloud with receipts

the verified data hub — every number recomputable

learn · how agents use websites

What is WebMCP, and what can an agent buy with it?

WebMCP is a proposed web standard that lets a web page give an AI agent a short list of named actions. The agent does not have to read the page and guess where to click. The Web Machine Learning Community Group writes it, and the editors named on the spec work at Google and Microsoft. As of 2026-10-06, it is a draft, not a standard.

Part of the Learn path. Every term links to the glossary the first time it appears.

What problem does WebMCP try to solve?

WebMCP gives a web page a way to tell an agent what it can do there. Without it, an agent can still use the page by acting like a person. Chrome's WebMCP page calls that approach "actuation."

"Actuation is the act of an agent simulating manual mouse clicks and text input, as though it were the human user engaging with your website."

Chrome's WebMCP page, last updated 2026-08-07

WebMCP works the other way round. Instead of the agent working the page like a person, the page lists the things an agent can do. The spec says the API lets web applications offer tools to AI agents.

Picture a shop counter. Without a menu, a customer has to look around and work out what is on offer. A menu on the counter lists it all.

How does a web page offer tools to an agent?

The page registers each tool, and an agent can then list the tools and run one. As of 2026-10-06, the draft does this through an object called document.modelContext. Earlier previews used navigator.modelContext. The spec moved it to the document in a change merged 2026-05-19.

The object has three methods. registerTool(...) adds a tool. getTools(...) lists the tools. executeTool(...) runs one.

How does a page register a tool in script?

A tool has four parts: a name, a description, an input schema, and a function called execute. The name, the description, and the function are required. The schema says what the tool needs to be told before it runs.

The spec marks the schema optional. Chrome's page says registration requires one. A name is 1 to 128 alphanumeric characters.

How does a form become a tool?

A second way starts from a form. The page adds toolname and tooldescription to a <form>. Form fields can add toolparamdescription.

This version comes from the explainer and from Chrome's docs, last updated 2026-09-25. It does not come from the spec's own text. The next section covers who presses submit.

What would OTTO do on a checkout page?

OTTO, our example agent, is an invoicing agent that buys OCR (text pulled out of scanned pages) at $0.01 a page. OTTO is invented for teaching. It describes nothing but OTTO.

Say OTTO reaches an OCR seller's checkout page. Say the page offers tools named like the examples Stripe lists for its own checkout pages. Each time OTTO runs a tool, that is one tool call.

  1. OTTO asks the page for its list of tools.
  2. OTTO runs get_order_summary. The page answers with the order: OCR processing at $0.01 a page.
  3. OTTO runs select_payment_method, then fill_payment_form, to set up the payment.
  4. OTTO is about to run submit_payment. That step would send the payment.

That is the pattern: list the tools, run them in order, and reach the last one. Before OTTO runs it, one question comes up. Who says yes?

Who decides before money moves?

The WebMCP draft does not decide. It lets a page attach hints to a tool, and the hints are advice. Nothing in the draft requires a confirmation prompt. Shopify and Stripe each tell agents, in their own docs, to get the user's go-ahead first.

What do the three hints say?

The spec defines three hints. Each is true or false, and each defaults to false. This is the spec's wording:

The spec says consequentialHint "acts as a signal to the client or agent." It is advice, not a rule. Chrome's own example tool is named book_flight, and it sets consequentialHint: true.

Think of a sticky note on a purchase order: "Big purchase. Ask first." The note does not stop anyone by itself.

What do Shopify and Stripe tell agents to do?

As of 2026-10-06, both docs pages carry no date. Shopify's docs say: "Before you call complete_checkout, show the buyer the current order and total, and get their permission to place it."

Stripe's docs say: "Stripe sets the WebMCP consequentialHint option for payment submission tools. Before calling a tool with this hint, get the user's confirmation. Don't interpret the hint itself as confirmation."

Who presses submit on a form tool?

By default, the person does. Without toolautosubmit, the person reviews the form and presses submit. A page can opt in to toolautosubmit. Then the agent can submit after filling the form "without requiring the user to check it manually before submitting."

Chrome's page also sets a limit on the design: "This API is primarily designed for local browser workflows with a human in the loop."

Which browsers have it?

As of 2026-10-06, no primary source says WebMCP is on by default in any stable browser. Chrome and Edge each run an origin trial, which is a limited test, not a full launch. Firefox's position is "neutral." WebKit's position, for Safari, is "oppose."

The browsers are listed in alphabetical order. The order is not a ranking.

Who has switched it on for shopping?

As of 2026-10-06, this page covers three dated announcements from two companies. Shopify switched on storefront tools on 2026-08-05 and checkout tools on 2026-09-28. Stripe switched on Checkout tools on 2026-09-22. They are listed in date order, which is not a ranking.

What did Shopify switch on for storefronts on 2026-08-05?

Shopify put WebMCP tools on what it calls every Liquid storefront and on Hydrogen preview storefronts. The tools let an agent search the catalog, browse, and get a product. They also let it get, update, or cancel the cart, proceed to checkout, manage orders, and search policies and FAQs.

Shopify says merchants need no setup. It also says agent support is "limited to Chromium-based browsers."

What did Stripe switch on for Checkout on 2026-09-22?

In a blog post by Cara Mecozzi and Steve Kaliski, Stripe wrote: "We enabled WebMCP on all Stripe Checkout UIs." Stripe lists four kinds: the full page, the embedded form, Elements, and the hosted invoice page. Sellers need no integration change.

Stripe's example tools are get_order_summary, select_payment_method, fill_payment_form, and submit_payment. In plain words, they read the order, pick how to pay, fill in the payment form, and submit the payment. Stripe's docs add: "WebMCP is an experimental browser capability based on a proposed web standard. Browser support, tool availability, and tool schemas can change."

What did Shopify switch on for checkout on 2026-09-28?

Shopify's changelog says: "Checkout WebMCP lets AI agents in the buyer's browser read and update the checkout, and place the order after the buyer confirms it." The tools are get_checkout, update_checkout, complete_checkout, and navigate_to_storefront. Shopify says they "don't expose a new API or require merchant configuration."

Shopify's docs say the tools are not available on several checkout types. Examples are B2B checkout, embedded and mobile SDK checkouts, and draft orders.

How is WebMCP different from MCP?

MCP is short for Model Context Protocol. An MCP server runs on a company's backend, and an agent connects to it over the network. WebMCP puts the tools in the page's own script, in the browser tab a person has open. There, they run.

The spec draws the same contrast. It says pages that use WebMCP "can be thought of as Model Context Protocol servers that implement tools in client-side script instead of on the backend." Client-side script is code that runs in your browser.

An everyday comparison: an MCP server is like an order desk you phone. WebMCP is like the menu on the counter of the shop you are standing in.

As of a commit dated 2026-09-28, the spec marks its reference to MCP as informative. That means it is background, not a requirement.

What does that mean for a person? The tools work inside the browser tab you have open. An agent has to visit a page to learn what tools it offers. Chrome says: "Clients and browsers must visit a site directly to know if it has callable tools."

What can't WebMCP tell you?

It cannot tell you whether the draft is settled or how many purchases it has completed. It also cannot tell you what its only figures show, whether a page needs it, or whether a seller is worth paying. Each limit is below, with its source.

Is the draft settled?

No. The spec's status text says: "It is not a W3C Standard nor is it on the W3C Standards Track." The document is a Draft Community Group Report. It is rebuilt often, and its latest commit is dated 2026-09-30.

The spec's section on forms reads: "This section is entirely a TODO." In plain words, that section is not written yet.

How many purchases have gone through WebMCP?

As of 2026-10-06, we found no independent count of purchases completed through WebMCP. No source we read publishes one.

What do the only published figures show?

The only figures come from Stripe's own test. Stripe's blog post of 2026-09-22 reports 42% fewer tokens and 38% fewer tool calls than a baseline. The 42% counts tokens, the chunks of text a model reads and writes. The 38% counts tool calls.

The test used 60 runs across six models, on the checkout of one fictional retailer. It is Stripe's own evaluation. It is not field usage.

Does a page need WebMCP for an agent to buy?

Not by Stripe's docs. They tell agents: "If the tool required to complete a task isn't available, fall back to standard browser automation." So a missing tool does not end the attempt.

Does WebMCP say whether a seller is worth paying?

No. It says nothing about the seller behind a checkout page. That is a different question, and other pages on this site take it up.

Keep going

This page covered what WebMCP is, which browsers have it, and who has switched it on for shopping. The pages below cover how an agent buys and how its money moves. One is a Pulse issue about checkout pages that get agent tools.