demo · v146

Tool call replay

A staged conversation: a fictional in-page assistant takes natural-language prompts and selects a tool from the page's registered WebMCP set, runs it, and gets a response. Pick a scripted scenario or freestyle. The trace shows the prompt → tool match → arguments → response → side effect.

Behind a flag: WebMCP / navigator.modelContext ships behind chrome://flags/#enable-experimental-web-platform-features in Chrome 146. The page registers a stub registry and matches prompts against tool descriptions so the demo works regardless.

scripts

or freestyle

tool registry

trace

0prompts
0tool calls
0items in cart
// 146 surface:
navigator.modelContext.registerTool({
  name: "addToCart",
  description: "Add one or more items to the cart",
  parameters: { type: "object", properties: { sku: {type: "string"}, qty: {type: "number"} } },
  invoke: ({ sku, qty }) => cart.add(sku, qty),
});
// A user agent (or another tool consumer) calls invoke() over the registered set.

see also