A page registers each capability as a tool with a name, a natural-language description, and a JSON schema for inputs and outputs, typically through a browser API called navigator.modelContext; the tool handler is ordinary page JavaScript that runs in the user's existing authenticated session. WebMCP is the browser-scoped counterpart to MCP, which instead connects an agent to an external server you host, and it grew out of the MCP-B library by Alex Nahas. It is a W3C Web Machine Learning Community Group work item with early browser previews behind flags, not a shipped platform primitive. OpenAI ran a WebMCP Challenge on Devpost with a September 3, 2026 submission deadline.