Agent runtime

WebMCP

A proposed web standard that lets a page expose MCP-style tools to agents running in the browser — navigator.modelContext — so a site becomes callable by agents, not just readable.

Operating principle

Production AI is a system: context, tools, permissions, traces, evals, and feedback loops around the model.

What it is

WebMCP is a W3C community proposal (driven by Google and Microsoft engineers, 2025) that gives web pages a browser API — navigator.modelContext — for registering typed, callable tools. An agent operating in the browser, whether an extension or a built-in assistant, discovers those tools and calls them with the user's consent. It complements server-side MCP: same tool-calling model, delivered by the page itself.

Why it matters

Most agent-web interaction today is scraping or brittle DOM automation, where the site has no say in what an agent does. A first-party tool surface flips that: the site author declares exactly what agents can do — search this catalog, check this price, submit this form — with schemas and consent built in. The spec is early, and publishing tools now is a cheap position on the agentic web.

How we use it

Public sites in our stack — this one included — register in-page tools such as page search and contact actions alongside a hosted MCP server, published Agent Skills indexes, and llms.txt. Four complementary surfaces, one goal: an agent that lands on the product can find out what it does and act without scraping.

Related resources