Updated
Updated What changed on
Fact-checked ChatGPT Atlas's retirement and how AI Mode handles UK restaurant bookings were re-checked against the 14 October 2026 source checks. WebMCP's browser status was re-checked on 14 October 2026 against the Chrome developer docs, Chrome Status, the Microsoft Edge origin trial page, the specification and its implementation status list, and the Mozilla and WebKit standards position threads. The Chrome enable steps were re-checked on Chrome's WebMCP developer docs the same day. Nothing else on the page was re-checked.
- Removed ChatGPT Atlas as a current in-browser agent from the text and the diagram: OpenAI retired its ChatGPT Atlas browser in August 2026.
- Corrected the AI Mode booking example: in the UK, AI Mode finds restaurant availability and links you to partners such as OpenTable or TheFork to finish the booking, rather than booking through OpenTable alone.
- Replaced the forecast that wider browser support would arrive in mid-to-late 2026 with where WebMCP stands now: origin trials in Chrome 149 and Edge 150, experimental support in Brave's Leo assistant, a neutral position from Mozilla and opposition from WebKit. Opera is no longer named, as the implementation status list published with the specification does not include it.
- Updated the steps to try WebMCP: Chrome's docs now use the chrome://flags/#enable-webmcp-testing flag for local testing and the origin trial from Chrome 149 for real visitors, replacing the Chrome 146 Canary instructions.
What is WebMCP and should you care yet?
The 30-second answer
WebMCP (Web Model Context Protocol) is a browser API co-authored by Google and Microsoft, incubated at the W3C Web Machine Learning community group. It lets a webpage expose its features (search, login, add-to-cart, submit-form) as structured tools that browser-based AI agents can call directly, instead of guessing where to click. Google shipped an early preview in Chrome 146 Canary on 19 February 2026. As of October 2026 it runs as an origin trial (a time-limited test that sites sign up for) in Chrome 149 and Edge 150, and Brave has experimental support in its Leo assistant. Safari's engine team has said it opposes the proposal. Honest position for marketing-only sites in 2026: monitor adoption, do not implement yet. For sites with forms, search, booking flows or any action layer worth exposing, the declarative API is cheap enough to start experimenting.
The 60-second technical version
WebMCP is a browser-side API co-authored by Google's Chrome team and Microsoft's Edge team, incubated through the W3C Web Machine Learning community group. It rides on top of the underlying Model Context Protocol that Anthropic open-sourced in November 2024, but the WebMCP layer itself is the Google + Microsoft work and lives inside the browser tab, not on a backend server.
The page declares what it can do (a Tool Contract). The agent in the same browser tab reads that contract and calls the tools directly. No DOM screen-scraping, no clicking through forms blind. The user asks their agent for something; the agent picks the right tool the site has exposed; the tool runs inside the user's logged-in session and returns structured data.
Why anyone is talking about it
Two reasons. First, agentic AI is the next interface layer. Google AI Mode's agentic features signal that agents acting on websites is the direction of travel. In the UK, AI Mode finds restaurant availability and links you to partners such as OpenTable or TheFork to finish the booking (April 2026). Not every agent has lasted: OpenAI retired its ChatGPT Atlas browser in August 2026. Second, schema and llms.txt only expose content. WebMCP exposes actions. A travel site can let an agent search and book; a SaaS site can let an agent query usage data; an e-commerce site can let an agent add to cart, all without the agent having to figure out where the buttons are.
The two ways to implement it
Top-ranking explainers (Webfuse, Codely, Semrush) all converge on the same split. Two paths, pick by complexity:
- Declarative API (HTML attributes). Add
toolname="login"andtooldescription="Log in to your account"to an existing form. The browser registers that form as a callable tool. Lowest-friction path. If your forms are already clean and well-labelled, this is most of the work done. - Imperative API (JavaScript). Use the new
navigator.modelContext.registerTool()browser interface to register tools programmatically. Useful when the action is not a simple form (filtering, multi-step flows, dynamic data). Heavier to wire up.
A concrete example: the login form
Plain HTML form today:
<form action="/login" method="post">
<input name="email" type="email">
<input name="password" type="password">
<button type="submit">Log in</button>
</form>
WebMCP-declared version:
<form action="/login" method="post"
toolname="login"
tooldescription="Log in to your account">
<input name="email" type="email">
<input name="password" type="password">
<button type="submit">Log in</button>
</form>
That is it. Two attributes. The agent reads the form's existing input types and required fields as the schema. The tool inherits the user's browser session, so password managers, autofill and cookies all still work the way they do for the human user.
How to try it in Chrome today
Chrome first shipped WebMCP behind a flag in Chrome 146 Canary in February 2026. Chrome's developer docs (last updated 7 October 2026) now give two routes. To test locally, go to chrome://flags/#enable-webmcp-testing, set it to Enabled and relaunch Chrome. To reach real visitors, join the WebMCP origin trial, which runs from Chrome 149. Install the Model Context Tool Inspector extension to see which tools a page has declared. Test against your own site by adding toolname attributes to a form, then asking the in-browser agent to perform the action.
What WebMCP is not
- Not a replacement for schema. Schema describes content. WebMCP describes actions.
- Not a replacement for llms.txt. llms.txt curates content for AI consumption. WebMCP exposes tools for AI use.
- Not the same as Anthropic's MCP. Anthropic's MCP is the underlying protocol and runs server-side (Claude desktop talking to your filesystem, Cursor talking to GitHub). WebMCP is browser-side only, inside the user's tab, on a webpage they have already loaded.
- Not a finished standard. Early preview in Chrome 146 Canary (Feb 2026), then origin trials in Chrome 149 and Edge 150. The specification is still a W3C community group draft, WebKit has said it opposes it and Mozilla's position is neutral.
- Not useful for content-only marketing sites. No action to expose means nothing for WebMCP to do.
Honest framing: WebMCP may become important within 18-24 months or may stall. For marketing-only sites in 2026 it is not yet worth implementing. For sites with API products, search functionality, booking flows or interactive services that AI agents could meaningfully use, early adoption gives a first-mover advantage. Pick based on whether your site has an action layer worth exposing.
Who should consider WebMCP in 2026
- Travel and hospitality (booking, search, reserve)
- E-commerce (search, compare, add-to-cart)
- SaaS with user-facing data (usage queries, account info, integrations)
- News and reference sites with structured data (specific article lookup, fact retrieval)
- Directory and listings sites (filtered search)
Who should skip WebMCP for now
- Consultancy and professional services sites
- Blogs and content marketing sites
- Brochure sites with no interactive layer
- Any site that has not yet shipped basic schema and llms.txt
Primary sources
- Chrome developer blog - WebMCP early preview announcement
- W3C WML community group - WebMCP GitHub repo (specification + issues)
- Forbes - Google Ships WebMCP, The Browser-Based Backbone For The Agentic Web (19 Feb 2026)
- Webfuse - WebMCP practical guide (longest treatment of the protocol mechanics in 2026)
- Semrush - WebMCP for marketers and SEOs (visibility-focused angle)
- dev.to - Chrome's WebMCP early preview walkthrough
More on WebMCP.
What does WebMCP stand for?
WebMCP stands for Web Model Context Protocol. It is a browser API co-authored by Google and Microsoft, incubated at the W3C Web Machine Learning community group. It lets a webpage declare its features as structured tools that browser-based AI agents can call directly. It rides on the underlying Model Context Protocol that Anthropic open-sourced in November 2024, but the WebMCP layer itself is the Google + Microsoft work.
Is WebMCP an official standard?
Not yet, but it has shipped as an early preview. Chrome 146 Canary released the early preview behind a flag on 19 February 2026 (Forbes covered the shipping news the same day). The standard itself is incubated at the W3C Web Machine Learning community group. As of October 2026 the specification is a W3C community group draft, which states that it is not a W3C Standard nor on the W3C Standards Track. Chrome has run an origin trial since Chrome 149, Edge has its own origin trial live in Edge 150, and Brave has experimental support in its Leo assistant behind a flag. Mozilla's standards position is neutral and WebKit, the engine behind Safari, has said it opposes the proposal.
How do I enable WebMCP in Chrome today?
For testing on your own machine, open Chrome, go to chrome://flags/#enable-webmcp-testing, set it to Enabled and relaunch the browser. For real visitors, your site needs to join the WebMCP origin trial, which runs from Chrome 149. Install the Model Context Tool Inspector extension to see which tools a page has declared. Add toolname attributes to a form on your own site and ask the in-browser agent to perform the action. Full steps are in Chrome's WebMCP developer docs.
Should I implement WebMCP on my site now?
Depends on what your site does. If you have forms (login, search, contact, checkout) the declarative API is cheap enough to experiment with: two HTML attributes per form. If you have a SaaS dashboard or a multi-step interactive flow, the imperative API is worth a pilot on one workflow. If you run a content-only marketing site, blog or brochure site with no real action layer, there is nothing to expose and the work has no payback yet. Wait until adoption broadens before spending engineering time on it.
How is WebMCP different from MCP?
Two different layers, similar names. Anthropic's MCP is the underlying protocol, runs server-side, used by desktop apps like Claude desktop and Cursor to talk to backend tools, file systems and databases. WebMCP is the Google + Microsoft browser-side adaptation, runs inside the user's tab, lets the in-browser agent call tools the page has declared. Both can coexist on the same product, doing different jobs.
Will WebMCP affect SEO?
Indirectly, if it gains adoption. It does not replace traditional SEO signals or schema. It adds an action layer that agents can call. The likely shape: sites whose tools are easy for an agent to use will absorb more agentic traffic; sites whose tools are not declared will get skipped for competitors that have done the work. None of this is confirmed yet at scale.
Jason Burns
Independent UK SEO, GEO and AI consultant. In search since 2007. Portfolio includes 3M, BlackRock, Unilever and E.ON. Owner of SEO Moves Ltd since 2014.
More about Jason →Unsure whether your site has a WebMCP use case?
Send the URL. I look at the site, work out whether you have an action layer worth exposing, and tell you whether the work would actually pay off. Quick answer, often "not yet".
Send the URL