Check this page with an assistantOpens a chat asking it to summarise this article and name the evidence behind each claim.
Claude opens with the prompt on your clipboard: Anthropic does not support prefilled prompts on the web, and we would rather copy it than ship a button that drops it.
Scraping Is an Agent Guessing at Your Interface
Today an agent that needs to do something on your site reads the rendered page, infers which element is the search box, guesses which button submits, and hopes the flow it half-recognises behaves the way similar flows usually do. It works often enough to be useful and fails in ways nobody can debug. WebMCP replaces the guessing with a declaration: here are the things you can do, here is what each one needs.
The specification came out of engineering work at Google and Microsoft, incubated through the W3C's Web Machine Learning community group, and surfaces to developers as navigator.modelContext. It has two halves. A declarative API turns ordinary HTML forms and page elements into callable capabilities with minimal markup, and an imperative JavaScript API lets a site register richer, dynamic tools for cases the markup cannot express.
The shift worth noticing is about control rather than convenience. Under scraping, the agent decides what it can do with your site and you find out afterwards. Under WebMCP, you decide what is offered and the agent works within it. For anyone who has watched automated traffic take actions on their site that nobody designed for, that inversion is the interesting part, not the developer ergonomics.
Where It Actually Is, as of Now
WebMCP began behind a testing flag in Chrome Canary and moved in June 2026 to an origin trial in Chrome 149, meaning a site can register a token and expose tools to browser agents on live production traffic. Google has named companies including Expedia, Booking.com, Shopify, Etsy, Instacart, Target, Redfin, Credit Karma, and TurboTax as experimenting with it.
Read that list carefully, because it is the clearest signal in this whole story. Every company on it operates a site where an agent completing a task, booking a room, reordering groceries, checking an account, has obvious value. None of them is there to improve search visibility. The people closest to the standard are treating it as transactional infrastructure, which is what it is.
Two caveats on status. An origin trial is a time-boxed experiment that can be extended, changed, or withdrawn, and browser history is full of proposals that got this far and no further. And Microsoft co-authoring the specification makes Edge support plausible rather than confirmed. What you can reasonably conclude is that the two companies who would have to implement it are the ones writing it, which is exactly the property llms.txt has always lacked.
The Category Error to Avoid
WebMCP has nothing to do with being found. It does not affect indexing, retrieval, citation, or ranking. It governs what an agent can do once it is already on your page inside a browser. Anyone marketing it as an AI visibility tactic has either not read the specification or is counting on you not having.
Expect that marketing anyway, because the pattern is well established. A technical proposal appears, its name enters the discourse, and within a quarter it is a line item on proposals from agencies who cannot explain what it does. That is precisely what happened with llms.txt, where a reasonable idea became a billable deliverable long before any evidence emerged that anything consumed the file.
The comparison is instructive rather than dismissive. llms.txt has a publication problem: tens of thousands of sites emit it and no major provider has committed to reading it. WebMCP has the opposite profile, with implementation coming from the browser vendors themselves and adoption being tested by companies with a commercial reason to care. That does not guarantee it survives. It does mean the two deserve different levels of attention, and neither deserves a place in a visibility strategy.
Who Should Care, and When
The value of declaring a tool is proportional to the value of the action. If an agent completing something on your site saves a customer real effort, WebMCP is worth understanding now. If your site explains a service and collects enquiries, there is no action to declare and nothing to gain.
| Site type | Is there an action worth declaring? | Reasonable posture now |
|---|---|---|
| Booking, ordering, or reservations | Yes, the core transaction | Prototype during the origin trial |
| Ecommerce with a real catalogue | Yes, search and cart operations | Do the product data work first, watch this second |
| Account or self-service portals | Yes, status checks and routine changes | Watch, and think about authorisation early |
| Publishers and content sites | Search, at most | Ignore for now |
| Service businesses and local sites | No meaningful action exists | Ignore entirely; the leverage is in being legible locally |
If you are in the top three rows, the useful preparation is not writing WebMCP code. It is knowing which operations you would be willing to expose, what each one requires, and what could go wrong if it were called a thousand times an hour by software rather than twelve times by people. Those answers are needed whatever the standard ends up being called.
The Questions That Outlive the Standard
Authorisation, rate limits, and reversibility. Every agent-facing capability is an interface with different assumptions about volume, intent, and accountability, and those three questions determine whether exposing it is a feature or an incident.
- Who is acting, and on whose authority?
An agent operating for a signed-in customer is a different actor from one operating on its own initiative. If your permission model cannot express that difference, adding agent capabilities is adding an authorisation problem before you have an authorisation model.
- What happens at machine volume?
A form that a human submits occasionally can be called continuously. Anything that costs money, sends a message, consumes inventory, or triggers a person needs limits designed for software rather than for the traffic pattern you have today.
- Can the action be undone?
Delegated actions get disputed more than direct ones, because the person did not personally make the choice. Reversibility stops being a nicety and becomes the difference between a support ticket and a chargeback.
- Which crawlers can even reach the page?
Agent traffic that arrives as a user-triggered fetch will be turned away by a blanket AI block, so a policy written to object to training data can quietly refuse a customer. Separate those decisions using the crawler decision guide.
- Does the page work without JavaScript running?
Declared tools sit on top of a page that still has to render and be indexed by conventional crawlers. A site whose content only exists after client-side execution has a more urgent problem than agent readiness, covered in the JavaScript SEO guide.
Questions People Ask About WebMCP
- What is WebMCP?
WebMCP is a proposed browser API that lets a website declare structured tools an AI agent can call, rather than leaving the agent to interpret the page visually and click through it. It was developed by engineers at Google and Microsoft and incubated through the W3C's Web Machine Learning community group, and it exposes those tools through a new browser interface called navigator.modelContext.
- Is WebMCP available to use now?
It is available to experiment with, not to rely on. It first appeared behind a flag in Chrome Canary, and in June 2026 it moved to an origin trial in Chrome 149, which means a site can register a trial token and expose tools on real production traffic. An origin trial is a time-limited experiment that can change or be withdrawn, so it is a place to learn rather than a platform to build a business process on.
- Will WebMCP help my site rank or get cited by AI?
No, and this is the most common misunderstanding about it. WebMCP governs what an agent can do once it is already on your page in a browser. It has no bearing on whether search engines index you, whether assistants retrieve you, or whether you get cited in a generated answer. Treating it as a visibility tactic is a category error, and any vendor selling it as one is worth avoiding.
- What is the difference between WebMCP and llms.txt?
Both propose a machine-friendly layer over a website, and their status is very different. llms.txt is a proposal with no confirmed consumer, and log studies show AI crawlers largely do not request it. WebMCP is being implemented in a browser by its authors, with a live origin trial and named companies testing it. One is a file waiting for adoption, the other is an API being adopted by the party that has to implement it.
- Should a small business implement WebMCP today?
Almost certainly not yet. The cost is engineering time against a specification that may change, and the benefit only exists for sites where an agent doing something on a user's behalf carries real value, such as booking, ordering, or account actions. If your site's job is to explain a service and collect an enquiry, there is nothing for an agent to operate and nothing to gain from declaring it.
