This site speaks WebMCP

WebMCP is an emerging web standard that lets a page hand real, callable tools to an AI agent browsing it - instead of making the agent scrape and guess. This site registers two.

An agent reading my homepage can already see the words. What it can't do from HTML alone is act - so this site uses the proposed navigator.modelContext API (a W3C incubation, shipping in Chrome behind an origin trial) to register tools an agent-enabled browser can call directly:

get_site_overview

Registered on every page - structured JSON describing who I am, what's here, and how to reach me. Orientation without scraping.

run_literal_machine

Registered on The Literal Machine - an agent can actually play the activity: submit PB&J instructions, get the robot's execution log and verdict back as structured data. When it does, the run streams into the visible page too.

How it's built

Zero dependencies - the API is small enough that a polyfill would obscure more than it helps. The site feature-detects the API and stays silent where it's absent. Site-wide tools register once per visit; the activity tool is owned by its page's Stimulus controller, registering on connect and unregistering on disconnect, which is the correct lifecycle under Turbo navigation (the page changes, the browser never reloads). The activity tool goes through the same token and rate-limit path as the human UI - an agent gets no side door.

Try it

In an agent-enabled browser (Chrome with the WebMCP origin trial or flag enabled, or any browser with the MCP-B bridge extension), connect an agent to this site and ask it what tools the page offers - then have it make the robot a sandwich. It will discover the specification lesson the same way my students do.

Your browser, right now

checking for the WebMCP API…