01
Agent-to-site communication, made explicit.
Today, agents screen-scrape and guess. WebMCP lets a site say plainly: "these are my tools, here's what they take, here's what they return." No fragile selectors.
A WEB MCP WORKSHOP
02 · The story
Sites have to speak two languages now — to the person at the screen, and to the AI agent helping them. WebMCP is the proposed standard that lets a website declare its own tools directly to an agent. One site, two audiences.
01
Today, agents screen-scrape and guess. WebMCP lets a site say plainly: "these are my tools, here's what they take, here's what they return." No fragile selectors.
02
navigator.modelContext.A page registers tools via registerTool({...}). The browser surfaces them to whichever agent the user trusts — Claude, Gemini, an extension. The page never knows which.
03
The same screens we design for humans need to be legible to agents. Naming a tool, structuring its inputs, telling the user when an agent acted — those are design choices.
See it live →
Beancrest coffee shop demoMocks the platform-owned ARIA floor I'm proposing in W3C Issue #65.
03 · Set up & build
A
Open chrome://flags, search WebMCP, enable the WebMCP for testing flag, relaunch.
Native runtime. No extension required. Recommended path.
B
On stable Chrome: install the MCP-B extension. It picks up any tools your tab registers.
Backup path if you can't run Canary.
C
The chip at the top of the builder below should turn green when WebMCP is detected. If it's grey, finish A or B first.
01Describe
No JS to write — your tool's handler is stubbed. The agent gets your inputs echoed with this description.
02Preview
04 · The artifacts
Every registered tool with a downloaded artifact lands here. Day-of: fills up as attendees ship. After May 30: it's the time capsule of what got built.
Your tool here
Build above, then add to wall05 · The room
Settle-in intros from May 30 — a snapshot of who was there, what they hoped for, and what they took home.