Concepts / agent → tools & data
WebMCP new
Web Model Context API · W3C Web Machine Learning Community Group; Google and Microsoft
- Spec
- W3C community draft · Chrome 146
- Published
- Reviewed
What it is
WebMCP is a proposed web standard that gives a page a way to say “here are the actions an agent can take on me”. The page registers tools on navigator.modelContext, and the browser presents them to whatever agent is driving it, in MCP’s tool format. The goal is to replace screen-scraping and brittle DOM automation with a contract the site controls.
How it works
There are two ways to register a tool:
- Declarative: annotate an existing HTML form with attributes. The browser derives a tool from the form’s fields.
- Imperative: call
navigator.modelContext.registerTool()from JavaScript with a name, description, JSON schema and a handler. This can react to app state.
Chrome 146 shipped the first implementation in February 2026. Other engines are in the working group but have not committed to dates.
When to use it
You run a website and want agents, including the browser’s own assistant, to complete tasks on it reliably. It is not a replacement for a server-side MCP server. WebMCP is for actions that live in the page, in the user’s logged-in session.
Watch out for
- One browser. Treat it as progressive enhancement.
- The spec is still a community-group draft. Method names have already changed once.
- A tool you expose runs with the user’s session. Scope it like you would a form.