WebMCP goes live, and your signup flow becomes something agents can call

WebMCP stopped being a browser demo in August, and the version of your product that AI agents interact with may now be written by someone other than you. Within days of each other, OpenAI added Site tools to the browser built into its ChatGPT desktop app, Shopify switched WebMCP tools on across every Liquid storefront, and Cloudflare shipped a developer preview that injects a WebMCP bridge into any site it fronts, with no change at the origin.

The short version: instead of an agent looking at your page, guessing which button is the signup button and imitating a click, your page registers named tools with descriptions and JSON input schemas, and the agent calls them. Google's Chrome team, running an origin trial from Chrome 149, calls the alternative actuation and frames WebMCP as a progressive enhancement rather than a rebuild.

What actually shipped, and what did not

Three things landed at once, which is why the story changed. OpenAI's Site tools let ChatGPT Work and Codex call tools offered by the page in front of them; per Search Engine Journal, the feature requires GPT-5.6 Sol or Terra, is disabled on Luna, and is unavailable in Enterprise and Edu workspaces. Shopify exposes ten tools on every Liquid storefront, including a proceed_to_checkout tool that carries the cart to checkout without completing the purchase. Cloudflare's preview is a dashboard toggle: it adds one same-origin script tag to each HTML response at the edge, and every tool in the preview runs in the visitor's browser with no round trip to Cloudflare.

Agents no longer have to guess their way through a page built for humans.

That line is from Cloudflare's own launch post — vendor material, as are Chrome's docs. What has not shipped is consensus. The specification is a W3C Web Machine Learning Community Group draft, not on the W3C Standards Track. SEJ reports that WebKit opposes the proposal on API design, privacy and security grounds, and that Mozilla is neutral. Support today is Chromium plus OpenAI's desktop app. This is a bet, not a platform.

Why a marketer should care about a browser API

Because the tools are your funnel. Chrome's own use-case list opens with software support: an agent guided to the right form with the fields already filled from the conversation. Its examples include a submit_application tool for structured forms and a run_diagnostics tool buried on a settings page. Translate that to a B2B software site and the candidate list is obvious — start a trial, request a demo, filter pricing by seat count, search the docs, open a ticket.

First, a tool name and its description are copy. They may be the only copy an agent reads before deciding what your product does. Nobody owns that surface by default, and on hosted platforms it is already spoken for: Shopify chose the initial ten tools and their descriptions across the whole platform, and SEJ notes Shopify's published documentation does not say whether merchants can disable individual tools or edit their wording. Cloudflare's packs are Cloudflare's. If you are on either, the agent-facing description of your store or site may already exist, authored upstream.

Second, this is a conversion-layer story, not an AI-visibility one. WebMCP tools only exist once an agent has the page open. They do not touch ranking, indexing or citations, and none of the documentation claims they do. If you have been treating "AI search" as one project, split it: getting cited is a discovery problem, and getting acted on is a different one that starts after the click.

The measurement hole nobody has filled

Across the material SEJ reviewed, neither OpenAI nor Shopify published figures for tool calls, completed tasks, error rates, checkout starts, purchases or conversion change — and nothing explains what merchants see in analytics or how an agent-assisted sale gets attributed. Cloudflare Radar tracks detected WebMCP adoption across the domains it scans, which measures deployment, not outcomes.

So if an agent completes a trial signup inside a signed-in tab, your analytics most likely records an ordinary session. No channel, no campaign, no way to separate agent-completed conversions from human ones. Assume that gap exists before it starts mattering, not after.

The security posture matters too. Tools run inside the user's authenticated session — the appeal and the risk. Chrome's guidance flags malicious tool definitions and poisoned tool responses as prompt-injection vectors, and OpenAI treats site-provided definitions and results as untrusted, reviewing each call and confirming purchases, deletions, messages and permission changes with the user.

• Run curl -s https://yoursite.example | grep webmcp and find out whether tools are already being injected on your behalf.
• List the five actions that carry revenue on your site. Those are your tools, in priority order.
• Write the tool names and descriptions yourself, in the same review as your homepage copy. They are positioning, not config.
• Ask analytics how an agent-completed signup appears today. The honest answer is "as a normal session".
• Skip the origin trial if you have no agent traffic to learn from. Chromium plus one desktop app is a small room.

WebMCP is real enough to plan for and too early to build a quarter around. The cheap move is to decide now what your product would let an agent do, before a platform decides for you.