The Rise of WebMCP

WebMCP is poised to replace today’s “guess-and-scrape” web interaction for AI agents by letting websites expose structured, callable tools directly to the browser—cutting both token usage and engineering pain. Instead of an agent reading raw HTML, interpreting screenshots, or trying to infer which buttons to press, a page can register functions (for example, “search products” or “fill this form”) so the model can invoke them like tools. The practical payoff is straightforward: one tool call can collapse what used to take dozens of clicks, scrolls, and intermediate reasoning steps.

The core problem WebMCP targets is that current agent workflows treat websites like a foreign language. Whether agents rely on multimodal screenshots or direct DOM access, they still need to translate web content into something the model can act on. Screenshot-based approaches burn large amounts of tokens per image, while HTML/DOM approaches still require summarizing and filtering out irrelevant markup like paragraph tags and CSS. That translation overhead becomes a recurring cost and a recurring source of brittleness—especially for complex, dynamic pages.

WebMCP reframes the interaction model: each web page can act like an MCP (Model Context Protocol) surface for the agent. The agent asks what it can read, click, or fill in, and the site responds with structured capabilities and results. The concept has been discussed in academic work and earlier industry proposals, including collaboration between Microsoft and Google on a spec for how to make this work without forcing every capability behind a traditional backend API. Normal human browsing remains intact; the change is about giving agents a more efficient, standardized interface.

A key emphasis is human-in-the-loop coordination. Rather than pushing toward fully autonomous browsing, the design supports workflows where agents act on a user’s behalf while still handing control back when needed—such as when a site lacks the exact product a user wants and the system pauses for confirmation. This aligns with a three-pillar framing described at the Web AI Summit: context (helping the agent understand what matters beyond the current screen), capabilities (actions like form completion using available user knowledge), and coordination (managing the back-and-forth between user and agent).

Technically, WebMCP uses two main APIs. The declarative API leverages existing HTML forms, adding metadata like tool names and descriptions so agents can discover and use them. The imperative API targets more complex, dynamic interactions that may require JavaScript execution, defining a schema for richer tools. Both run client-side in the browser, which keeps the system closer to the user’s session and avoids forcing every site feature into a separate server API.

Chrome’s early preview already places WebMCP behind a flag, and participation in the Chrome early preview program is available for developers. The expectation is that rollout will accelerate soon, potentially alongside new ways to package site functionality into reusable “web MCPs,” making it easier for agent builders and product teams to ship reliable web actions with lower token costs.