AI agents interact with websites like tourists who don't speak the language. They take screenshots, they guess which button does what, they click and hope for the best. It works—barely—and it's slow, expensive, and hilariously unreliable. WebMCP changes that equation by letting your website declare its capabilities as structured tools any AI agent can discover and call directly. No screenshots. No guessing. Just function calls.
Two Flavors of WebMCP (And Why Most Articles Get This Wrong)
There are actually two different projects using the same name, and they're not compatible. The W3C/Chrome browser spec adds a navigator.modelContext object directly into the browser, letting websites register tools via JavaScript or plain HTML form attributes. Google has published official documentation on developer.chrome.com, confirmed an origin trial in Chrome 149, released sample apps in the GoogleChromeLabs/webmcp-tools repo, and even shipped a Model Context Tool Inspector Chrome extension for testing against gemini-3-flash-preview. This is the version you'll see covered by Forbes and VentureBeat—it's currently behind chrome://flags/#enable-webmcp-testing in Chrome Beta but heading to stable soon. The second option is Jason McGhee's open source JavaScript library (602 GitHub stars, MIT licensed) that works today in any browser with any MCP client. You add a script tag, register your tools, and users connect via a local WebSocket bridge using one-time tokens. No Chrome Beta required. The author of the implementation guide we're covering here used this version—because it actually ships now rather than waiting for Chrome's rollout timeline.
Implementation in 20 Minutes Flat
The author's step-by-step breakdown is refreshingly practical. He downloaded webmcp.js (29KB, zero dependencies) from GitHub releases v0.1.5, created a dedicated /mcp/ page to host the connection widget so it wouldn't clutter his blog's reading experience, and registered four tools: search_posts for keyword searches against his API, get_post for fetching article content by slug, list_posts via RSS parsing (always up-to-date), and get_site_info for giving agents context about who he is. The color theme pulled from his existing CSS custom properties so it matched both light and dark modes automatically. The connection flow is straightforward: users generate a WebMCP token from their MCP client (Claude Desktop, Cursor, etc.), paste it into the widget on the site, and they're connected. No API keys exchanged. Token gets discarded after registration. Everything runs client-side. For e-commerce specifically, he imagines search_products, check_price, view_inventory, add_to_cart tools—imagine telling Claude to find you a blue running shoe under £100 from your store and it just calls the right functions without touching a single screenshot.
The Agent SEO Angle Nobody's Talking About Yet
This is where things get interesting for site operators. AI agents are becoming primary web users—not just developers testing tools, but regular people delegating tasks to their assistants. WebMCP opens up optimization surfaces that don't exist yet for most sites. Dan Petrovic at Dejan frames it as four layers paralleling classic SEO: tool discoverability (the new indexing problem), tool descriptions (conversion copy for LLMs choosing which function to invoke), schema design (structured data's successor), and agentic CRO (A/B testing tool descriptions, watching success rates). His sharpest point: sites with well-structured WebMCP tools will capture agent-driven traffic; sites without them won't exist in the decision space at all. You won't be on the shortlist—you simply won't be considered.
The Unsolved Discovery Problem
The author acknowledges what's not worked out yet: discovery. Agents that already know about your site can use your tools, but nobody's telling new agents you have capabilities to begin with. There's no Googlebot equivalent for WebMCP—no registry, no directory, no index of tool-capable sites. His prediction: the SEO playbook repeats itself here. A directory pattern emerges first, a few become canonical, then platforms either acquire or replicate them. Within 12 months, expect an Ahrefs-style index tracking WebMCP-enabled sites and their tools. The play right now is to be one of the first ones in it.
Cloudflare Workers for the Infrastructure-Averse
For those who don't want to touch code themselves, Bastian Grimm built cf-webmcp—a Cloudflare Worker that injects WebMCP plus seven other discovery surfaces (manifest files, Link headers, llms.txt augmentation, AGENTS.md, api-catalog, Agent Skills) from a single TOML config. Three starter templates ship: Default, WordPress, WooCommerce. Live demo at webmcp.basgr.com. MIT licensed. If you're running WordPress or WooCommerce and can't modify the codebase directly, this is your fast path.
The Bottom Line
WebMCP isn't vaporware—the Chrome spec is shipping, the open source library works today, and the author proves you can be agent-ready in under an hour with zero backend changes. But the real opportunity is being early to a discovery index that doesn't exist yet. Every site that ships WebMCP tools now gets indexed before the directory pattern matures into something harder to break into.