WebMCP: A Standard for Websites to Expose Structured Tools to AI Agents
This article, reposted from the Agentic AI Foundation blog, describes WebMCP: developers can register named tools via document.modelContext.registerTool, giving each tool an input schema and an execute function, and the browser makes those tools available to a supported agent while the page is active; existing HTML forms can also expose tools declaratively through attributes such as toolname, tooldescription and toolparamdescription. It draws a distinction with MCP: an MCP server can offer tools when no related page is open, whereas a WebMCP tool is attached to the current document, can use the page's state and the person's session, and can be combined with MCP Apps and A2UI. The piece says experimental implementations exist — Chrome via a flag and origin trial, a Cloudflare developer preview, and Brave's testing with its Leo assistant — and that OpenAI reviews each site tool call and asks the person to confirm sensitive actions, while stressing that servers must still apply their existing access controls and input validation.
If browsers adopt it broadly, WebMCP could change how sites expose capabilities to agents, though the draft, permissions model and security boundaries are still evolving.