What Is WebMCP? The Next Layer of AI Discovery

Content is for reading. WebMCP is for doing.

llms.txt and structured data tell an AI system what a page is about. Neither tells an agent what a page lets you do. WebMCP closes that gap: a page declares its interactive tools (search, filter, add to cart, submit a form) the same way it declares a title or a schema, so an MCP-compatible agent can use them directly instead of guessing at a UI it was never built to see.

Two ways to declare a tool

WebMCP defines a browser API, navigator.modelContext.registerTool(), that a page calls to register a tool with a name, a description, and parameters — the imperative path. A simpler declarative path exists too: toolname and tooldescription attributes placed directly on the HTML element that triggers the action, no JavaScript API required.

• The imperative path calls registerTool() from JavaScript, typically inside a bundled module.
• The declarative path uses toolname/tooldescription attributes on the element itself, readable straight from raw HTML.

Why most sites test as “none” even when they have working tools

A real bug in our own engine illustrates how easy this is to get wrong. GEO Optimizer's WebMCP check originally looked for registerTool() only in the raw HTML response. That's invisible for almost any real implementation, because Astro, Vite, Next, and SvelteKit all emit that call inside a hashed, externally-loaded JS bundle, never inline. A site with five working tools, a published manifest, and a webmcp block in its own /ai/summary.json still scored readiness_level: none.

The fix wasn't to execute JavaScript on every audited page (too expensive to run at scale) or to fetch and parse every bundle (fragile, since there's no reliable way to know which file to look in). It was simpler: /ai/summary.json is data the audit already downloads for a different check, so reading its webmcp block, for example {"webmcp": {"available": true, "tools": [...]}}, costs zero extra requests and credits partial readiness when the static HTML scan finds nothing. If your tools are implemented behind bundled JS, declaring them in /ai/summary.json too is the one addition that makes them visible to any static scanner, not just this one.

How readiness is actually scored

The check looks at two groups of signals. WebMCP signals: registerTool() in raw HTML, toolname/tooldescription attributes, or a declared webmcp block in /ai/summary.json. Broader agent signals: schema.org potentialAction (SearchAction, BuyAction, OrderAction), forms with real labels an agent can parse, and a linked OpenAPI/Swagger spec.

• Advanced requires at least one WebMCP signal plus two or more agent signals.
• Ready needs either a WebMCP signal on its own, or two-plus agent signals without one.
• Basic means at least one agent signal, but no WebMCP signal.
• None means nothing detected.

This is diagnostic, not scored: it carries zero weight in the 100-point GEO score today. That's deliberate. The standard is still forming, and folding an early, largely optional signal into a score used to compare sites would penalize sites for not adopting something with no confirmed adoption curve yet.

Should you implement it now?

Depends on what your site actually does. If it's informational — a blog, a marketing page, a guide — there's nothing to declare; WebMCP describes actions, and you don't have any beyond navigation. If your site has real interactive surfaces — search, filtering, checkout, account actions, a dashboard — it's the first formal way to make those legible to an agent instead of leaving it to guess at a UI built for a mouse and a screen.

Our own roadmap lists WebMCP readiness as exploring, not shipped. The spec itself is early, with no confirmed adoption curve or citation impact yet. Treat this guide as a map of what exists today, not a claim that skipping it costs you anything right now.

Further reading

• What Is llms.txt? — the declarative content layer WebMCP's action layer sits next to.
• AI crawlers and robots.txt: the 2026 guide — the access layer underneath both.
• Structured data for AI citations — potentialAction and the other schema signals WebMCP readiness also checks.
• How Claude Cites Sources: What We Know So Far