WebMCP · navigator.modelContext
Your website has a second audience now — and it can't click
A growing share of your visitors are AI agents acting for real people. To them, your site is a wall of markup: they scrape the DOM, guess at buttons, and miss anything JavaScript renders late. WebMCP fixes that — and it's barely six months old.
What just changed
WebMCP — the W3C Web Model Context API, navigator.modelContext — lets a page hand
agents real tools instead of making them scrape. A tool is a named function with a
description and a schema ("search this catalogue", "book this slot") that an agent can call
directly, in the page, while the human watches.
The idea moved fast: proposed by Microsoft's Edge team in August 2025, co-authored by Google's Chrome team weeks later, accepted into a W3C community group that September, and shipped as a developer preview in Chrome in February 2026. It is early — which is exactly the point. Almost no site on the web registers tools yet. The ones that do are, for the moment, effectively the only shops on a brand-new high street.
"Agent-ready" is now measurable
The free AIEO scan scores any domain across four groups: Traditional SEO, AI Discoverability, Schema Coverage — and Agent Interaction (WebMCP), worth 20% of the overall score. It detects whether your pages register callable tools, and how many. Every failed check ships with the exact fix.
And because it felt wrong to score sites on agent-usability from a site agents can't use, this site registers its own WebMCP tools — six of them. Ask your agent "is my site agent-ready?" on the homepage and it runs the scan for you, with the score card rendering in front of you while a small pill names each tool call. The auditor eats its own cooking: it scores 100 on its own agent check.
A real before-and-after
Theory is cheap, so we instrumented a live production site: enm-compass.life, a relationship-style quiz — 33 questions, three axes, radar-chart results, all computed in the browser.
An agent can now take the quiz with you: you talk through your preferences in plain language and the answers sweep visibly across the page. The part we care most about — the site's privacy promise, your answers never leave the browser — survived untouched. WebMCP tools execute in-page, so the agent experience inherits the same privacy contract as the human one.
What we learned scanning the real web
One finding worth sharing with anyone adopting WebMCP: modern bundlers hide your
tools. Vite, Astro and webpack routinely split the code that calls
registerTool into a chunk that no <script src> tag ever names —
it's reachable only through an import statement inside another script. A naive
scanner (including our first version) under-counts almost every real-world build. The AIEO
scanner now follows same-site import chains, because that's how the web actually ships
JavaScript in 2026. If you're checking your own site by grepping the HTML for script tags,
you'll miss your own tools.
Try it with your own agent
Enable WebMCP in Chrome at chrome://flags/#enable-webmcp-testing (or use the MCP-B
extension), open the homepage, and ask your agent to scan a site. Prefer raw
data? GET /api/agent-scan?domain=example.com returns the full scan as JSON — CORS
open, built for agents.