Do you need WebMCP for AI agents to use your website?

No. Agents already interact with sites by reading the page and simulating clicks, and they will keep doing that on sites without WebMCP. What WebMCP changes is reliability: instead of the agent inferring your interface, you declare a typed function with a schema and the agent calls it. Think of it as an optimisation for flows where a misinterpretation is expensive, such as checkout, booking, or account changes. Clean semantic HTML and accurate structured data remain the foundation underneath it.

Which browsers support WebMCP today?

WebMCP is available in Chrome through an origin trial starting in Chrome 149, which means you can enable it for real users on your own origin while the API is still provisional. It is not a shipped, generally available web platform feature, and it is being developed as a proposal in the W3C Web Machine Learning Community Group with both Google and Microsoft contributing. Check the current Chrome documentation before planning a rollout, since the API surface has already changed once during the trial period.

Does WebMCP replace structured data and schema markup?

No, and treating it as a replacement would be a mistake. WebMCP tools only exist while a user has your page open in a browser tab, so they cannot help a crawler indexing your site, an assistant answering from training data, or a search engine building a result. Structured data and clean semantic HTML are how your facts travel without a live session. WebMCP handles the action layer for an agent already on your page. The two operate at different points in the journey.

Is WebMCP a security risk for logged-in users?

It introduces a surface worth designing carefully, which is why the specification includes gating. Chrome's documentation notes the APIs are restricted by the tools Permissions Policy, defaulting to the site's own origin, and that WebMCP is only available in origin-isolated documents, so embedded third parties do not get access by default. The practical guidance is the same as any privileged API: expose the narrowest tools you can, validate every parameter server-side, and require explicit confirmation for anything that moves money or changes account state.

WebMCP: Expose Your Website as Tools for Browser AI Agents (2026)

WebMCP is a proposed web standard that lets a site declare typed tools an in-browser AI agent can call directly, using the document.modelContext JavaScript API or annotated HTML forms. It entered an origin trial in Chrome 149. The goal is to replace brittle DOM scraping and simulated clicks with explicit, validated function calls.

The proposal is a joint Google and Microsoft effort developed through the W3C Web Machine Learning Community Group, where WebMCP is a chartered deliverable described as an API that lets developers expose their web application functionality as tools (W3C Web Machine Learning CG).

WebMCP replaces DOM scraping with explicit tool calls

The problem WebMCP solves is that agents currently guess at your interface. The explainer states the goal plainly: enable AI agents to be more reliable and helpful by interacting with websites through well-defined client-side tools instead of brittle UI actuation such as DOM scraping and simulated clicks (W3C Web ML CG explainer). Actuation is the term for an agent pretending to be a mouse.

Chrome's documentation frames the contrast as a single image: on the left the agent scrapes the DOM, on the right the agent calls a tool (Chrome for Developers). Multi-step actuation leaves every step open to interpretation, and each interpretation is a chance to get your checkout flow wrong. This is the next layer on the groundwork in optimizing your website for AI agents.

The WebMCP origin trial opened in Chrome 149

WebMCP is testable in production today under an origin trial. Chrome's announcement is direct: in Chrome 149, you can sign up for the WebMCP origin trial (Chrome for Developers). An origin trial means you can enable the API for real users on your own domain while the design is still changing.

The timeline ran across three milestones this year. Chrome published the WebMCP documentation on 18 May 2026, unveiled the agentic web vision at Google I/O on 19 May, and opened origin trial signup on 9 June (Chrome for Developers). Treat anything you build now as provisional, the same discipline we argue for in agentic coding.

The imperative API registers tools through document.modelContext

The imperative WebMCP API lives on document.modelContext, and getting this name right matters because it moved. The explainer describes an imperative API on the web platform under document.modelContext, built around registerTool and a toolchange event (W3C Web ML CG explainer), and Chrome's reference documents the companion getTools and executeTool methods (Chrome for Developers). Each tool carries a name, a description, and a JSON Schema for its parameters.

One deprecation will break early code. Chrome's imperative API reference notes that navigator.modelContext is deprecated in Chrome 150 and instructs developers to use document.modelContext instead (Chrome for Developers). Any tutorial written before that change is already wrong, which is a good reason to read the source rather than the summaries, as we argue in our Model Context Protocol explainer.

The declarative API turns existing HTML forms into agent tools

Most sites can expose their first WebMCP tool without writing JavaScript at all. The declarative API works through HTML attributes on a form: toolname, tooldescription, an optional toolparamdescription on individual fields, and toolautosubmit (Chrome for Developers). A search form or a support request form becomes a callable tool with two attributes.

The form lifecycle gains agent-aware hooks alongside it. Chrome documents SubmitEvent.agentInvoked so you can tell an agent submission from a human one, SubmitEvent.respondWith for returning a structured result, toolactivated and toolcancel window events, and the CSS pseudo-classes :tool-form-active and :tool-submit-active for visual state (Chrome for Developers). Knowing which submissions came from an agent is exactly the measurement problem behind B2B website conversion.

WebMCP and MCP solve genuinely different problems

WebMCP is not a browser port of the Model Context Protocol, and Chrome says so directly: WebMCP is not an extension or a replacement of MCP, and is better understood as a set of MCP-inspired APIs (Chrome for Developers). The naming has caused a lot of avoidable confusion.

The reasoning is architectural. The explainer notes MCP was built primarily for server-to-client and stdio process communication, and lacks native web concepts like origins, standard browser permissions, and DOM integration (W3C Web ML CG explainer). The W3C charter goes further, stating the WebMCP API is agnostic with respect to underlying protocols. Both matter to anyone planning a connected marketing and technology stack.

WebMCP permissions policy, origin isolation, and ephemeral tool lifetime

WebMCP ships with real constraints, and they are the first thing to check before you plan a rollout. Chrome documents that both APIs are gated by the tools Permissions Policy, which defaults to self and requires an allow attribute on an iframe, and that WebMCP is only available in origin-isolated documents (Chrome for Developers). Third-party widgets do not get your tools for free.

Tool lifetime is the constraint that shapes the use cases. WebMCP tools are ephemeral: they exist only while your page is open, and once the user navigates away or closes the tab, the agent cannot access your site (Chrome for Developers). Chrome also frames headless use as possible but outside the design intent, since the API targets local browser workflows with a human in the loop. Persistent, server-side discoverability still runs through markup, as covered in schema markup for AI search.

Preparing for the agentic web before WebMCP standardises

The pragmatic move now is to instrument and experiment rather than rebuild. Chrome DevTools ships a WebMCP pane that lets you inspect, debug, and test the tools your site exposes, including error messages when parameters or return values violate your declared schema (Chrome DevTools documentation). You can validate a tool before any agent calls it.

There is enough momentum to justify a small bet. The proposal is a synthesis of two proposals from Microsoft and Google (W3C Web ML CG minutes), and Google's I/O post says global consumer brands are already experimenting with WebMCP. Start with one high-value flow, expose it declaratively, and keep your content readable for the engines that will never call a tool at all, which is the discipline behind answer engine optimization. If you want help scoping that first tool, we can map it with you.

WebMCP: How to Expose Your Website as Tools for Browser AI Agents

WebMCP replaces DOM scraping with explicit tool calls

The WebMCP origin trial opened in Chrome 149

The imperative API registers tools through document.modelContext

The declarative API turns existing HTML forms into agent tools

WebMCP and MCP solve genuinely different problems

WebMCP permissions policy, origin isolation, and ephemeral tool lifetime

Preparing for the agentic web before WebMCP standardises

Frequently Asked Questions

WebMCP replaces DOM scraping with explicit tool calls

The WebMCP origin trial opened in Chrome 149

The imperative API registers tools through document.modelContext

The declarative API turns existing HTML forms into agent tools

WebMCP and MCP solve genuinely different problems

WebMCP permissions policy, origin isolation, and ephemeral tool lifetime

Preparing for the agentic web before WebMCP standardises

Frequently Asked Questions

Related Articles

How to Optimize Your Website for AI Agents in 2026

What Is the Model Context Protocol (MCP)?

Agentic Coding in 2026

Strategy, AI, and growth: delivered to your feed.