WebMCP (Web Model Context Protocol)

WebMCP lets a website tell an AI agent working in the browser what it can do on the page, such as search a catalogue or book a table, by publishing those actions as tools the agent can call. It is a proposed standard, co-developed by Google and Microsoft, that comes into play when an agent acts on a page rather than when a crawler indexes it, so it has no documented effect on rankings.¹² Its relevance is to sites where agents act rather than only read, such as e-commerce, travel and booking.

What is WebMCP?

MCP (Model Context Protocol) was introduced by Anthropic on 25 November 2024, open-sourced as “a new standard for connecting AI assistants to the systems where data lives”.³ It has since been adopted broadly across the AI industry. WebMCP is an adaptation of that protocol for the browser: a way for a website to declare its own callable tools in a format that a browser’s in-built AI agent can discover and invoke.

WebMCP (Web Model Context Protocol) entered early preview in February 2026. It was developed through the W3C Community Group with Google and Microsoft as co-developers, and became available in Chrome 146 Canary behind the “WebMCP for testing” flag, with Edge 147 adding equivalent flag-gated support in April 2026.⁴

Where a conventional web page communicates with users through HTML that humans read and navigate, a WebMCP-enabled page also communicates with AI agents through structured tool definitions that machines can call directly via the document.modelContext JavaScript API.¹

WebMCP introduces two APIs:

• Declarative API: annotations added to a standard HTML form turn it into a tool, with the form’s fields as the tool’s parameters
• Imperative API: dynamic interactions requiring JavaScript, for complex or stateful operations¹

WebMCP is not a crawler-facing protocol. Chrome describes it as a bridge between web applications and agents, and neither its WebMCP documentation nor Google’s AI optimisation guide connects it to indexing or ranking.¹² It operates at the browser and agent layer, at the moment a user’s AI agent takes action on their behalf.

How does WebMCP differ from structured data?

Structured data and WebMCP address different questions. Structured data says what a page contains. WebMCP says what can be done on it.

	Structured data	WebMCP
Question answered	What does this page contain?	What can be done on this page?
Consumed by	Search crawlers, indexing systems	In-browser AI agents at action time
Format	JSON-LD / schema markup	Tools the page registers, in JavaScript or as annotated HTML forms; an agent finds them only by visiting the page1
Example	This is a Product with a price of £49	Here is a checkAvailability function

The two are complementary. Google says structured data is not required for its generative AI features but keeps its rich-result eligibility;² Microsoft says structured content, with schema-marked pages as its example, helps its AI answers interpret and summarise a page.⁵ WebMCP enables AI agents to act on content on behalf of users. A site implementing both gives AI agents both the context to understand what they are looking at and the tools to take action on it.

How does WebMCP differ from agentic search?

Agentic search is about the content retrieval layer: structuring pages so that AI research agents can find, read, and synthesise relevant passages across a multi-step task. The agent browses, reads, and reports.

WebMCP is about the action layer: declaring what an agent can do on a site once it gets there. Instead of simulating user behaviour (navigating to a search box, typing a query, scrolling through results), an agent calls a searchProducts function and receives structured JSON back in a single operation.

A site without WebMCP can still be cited and retrieved by agents. WebMCP determines whether agents can also act on a site’s behalf, which is a different capability entirely.

Which sites should implement WebMCP?

The value of WebMCP scales with how much of a site’s utility involves completing tasks rather than reading information.

Highest value as agent clients adopt it:

• E-commerce: searchProducts, checkAvailability, addToCart, trackOrder are the obvious candidates. An agent shopping for a user needs to query inventory, check delivery options, and initiate a purchase. Without WebMCP the agent works by actuation, Chrome’s term for simulating clicks and typing, which Chrome says “may have numerous steps and leaves each step open to interpretation by the agent”; with it, a few structured calls.¹
• Travel and booking: checkDates, searchFlights, bookAppointment, cancelReservation. Agentic travel research is already a primary use case for tools like Gemini Spark.
• Local services: requestQuote, checkAvailability, bookService. Google’s agentic booking feature (announced at I/O 2026) handles calls to businesses for home services, beauty, and similar categories, though this runs on Google’s own infrastructure rather than WebMCP directly.⁶
• SaaS and tools: any site where the primary action is doing something in a logged-in context: creating a document, running a report, triggering a workflow.

Lower priority for now:

Editorial and informational sites. Agents read and synthesise this content but do not act on it. The relevant investment for these sites remains content quality and retrieval optimisation, not WebMCP implementation.

Current state and timeline

WebMCP has been available in Chrome 146 Canary behind the “WebMCP for testing” flag at chrome://flags since 10 February 2026. Edge 147 (April 2026) added the same flag-gated preview; in both browsers the flag must be enabled manually. At Google I/O in May 2026, Google announced a public origin trial in Chrome 149, live from approximately 2 June 2026.¹ That is the step between flag-behind testing and stable release. Those version numbers are history rather than the current state: Chrome stable has moved several releases past 149 since the trial opened, so check your own build rather than the milestone named here. Chrome’s announcement does not name an end milestone for the trial, and secondary coverage that does is not corroborated by Google, so treat the trial as open-ended until Chrome says otherwise.

The W3C Web Machine Learning Community Group added WebMCP as a formal deliverable in September 2025. Google and Microsoft are co-developers. Firefox and Safari have not announced plans to support WebMCP; cross-browser parity is a dependency for agent adoption at scale. Neither Chrome nor Edge has named a milestone for stable support, and Chrome’s WebMCP documentation gives no stable release date and no end date for the origin trial.¹

The API has changed name since the trial opened, and the migration is now complete. The W3C draft moved the getter from Navigator to Document in May 2026, on the reasoning that tools belong to a page rather than to the browser. navigator.modelContext was deprecated in Chrome 150 and kept only as a backward-compatible alias; Chrome’s own WebMCP documentation now uses document.modelContext throughout, with no deprecation notice left to read, because there is no longer a transition in progress. Anything still written against navigator.modelContext is working through an alias that Chrome has said it will remove.

As of June 2026, no major AI agent called these tools in production. That changed in August 2026, when OpenAI added WebMCP to the built-in browser of the ChatGPT desktop app, where it calls the tools “site tools”: they work when “your account has access and the website provides site tools”, not in Chrome, and ChatGPT still asks for confirmation before sensitive actions such as purchases.⁷ That browser is not ChatGPT Atlas, OpenAI’s standalone browser, which OpenAI set to stop working on 9 August 2026 as it moved browser-based agent work into the ChatGPT desktop app and Codex.⁸ Google said at I/O in May 2026 that Gemini in Chrome “will soon support WebMCP APIs”;⁹ Chrome’s documentation does not yet say it has. Chrome’s documentation still states that WebMCP “is under active discussion and subject to change in the future”.¹

On the site side, Shopify turned WebMCP on across its platform. Since 5 August 2026 every Liquid storefront, and the Hydrogen developer preview, has registered tools for catalogue search, product pages and the cart, with “nothing to install or configure”.¹⁰ Since 28 September 2026 eligible checkouts register four more, which let an agent fill in contact, delivery, discount and saved Shop Pay payment details and place the order once the buyer confirms. The standard three-page checkout registers them only when the buyer pays with Shop Pay, and B2B and embedded checkouts not at all.¹¹ Shopify’s documentation says agent support is “currently limited to Chromium-based browsers”.¹⁰

Gemini Spark, Google’s personal AI agent announced at I/O 2026, launched with MCP connections to Canva, OpenTable, and Instacart, and Google said more partners were “integrating now”.¹² MCP is becoming a common interface between AI agents and the services they use. Whether WebMCP follows depends on agent clients calling its tools, which none had done in production by June 2026 and ChatGPT’s desktop browser began doing in August.⁷

The spec continues to evolve. Treat current implementations as experimental.

The agent protocol stack

WebMCP is one layer in a broader stack of agent protocols. Each addresses a different coordination problem.

MCP (Model Context Protocol): the open standard developed by Anthropic and now adopted across the AI industry, including by Google and Microsoft. MCP defines how AI models connect to external tools and data sources. An MCP server built for one platform works with any MCP-compatible agent. Gemini Spark launched with MCP support and three named integrations (Canva, OpenTable, Instacart), with more partners “integrating now” in Google’s words; the summer partner names that circulated in coverage are not in Google’s announcement.¹²

WebMCP: the browser-native adaptation of MCP covered in this article. Where standard MCP connects server-to-server, WebMCP exposes callable tools directly in the browser via the document.modelContext API, at the moment a user’s in-browser agent takes action.

A2A (Agent-to-Agent Protocol): an open standard for coordination between agents, originally developed by Google and donated to the Linux Foundation in June 2025 for neutral governance. When one agent delegates a sub-task to another (for example, Gemini Spark handing a specialised research task to a retrieval agent), A2A manages the handoff, authentication, and result passing. Supported by more than 150 organisations as of April 2026, including Microsoft, AWS, Salesforce, SAP, Cisco, IBM and ServiceNow, with native integrations in Azure AI Foundry, Amazon Bedrock AgentCore and Google Cloud.¹³¹⁴ Publishers do not implement A2A directly; it operates at the infrastructure layer and explains how multi-agent chains resolve across different systems.

UCP (Universal Commerce Protocol): an open commerce-layer protocol co-developed by Shopify and Google for agent-initiated transactions, announced by Google CEO Sundar Pichai at the National Retail Federation conference on 11 January 2026 with backing from more than 20 retailers and payment networks.¹⁵ UCP exposes checkout endpoints that AI shopping agents can invoke directly to complete a purchase, rather than navigating a standard checkout flow.¹⁶ Covered in agentic commerce.

ARD (Agentic Resource Discovery): the discovery layer that sits in front of the others. Announced in June 2026 by Google, Microsoft, Hugging Face, and others, ARD defines a manifest at /.well-known/ard.json (renamed from ai-catalog.json in August 2026) that lists a site’s agentic resources (MCP servers, A2A agents, skills, APIs) so registries can index them and agents can find them by natural-language query.¹⁷¹⁸ Where WebMCP declares the tools, ARD is how an agent discovers that those tools exist. It is an early proposal with no confirmed adoption yet; see agent-readiness.

For most informational and editorial sites, MCP and WebMCP are the relevant layers: they cover what your site can offer agents that interact with it. A2A operates between agents and is not publisher-facing. UCP applies to commerce sites managing agent-initiated transactions, and ARD applies once a site has callable capabilities worth advertising.

Security and user consent

Two design decisions in the WebMCP spec matter for early implementers.

WebMCP tools are gated by the tools Permissions Policy, which defaults to self: tools can be registered in top-level and same-origin contexts but not in cross-origin iframes, unless the iframe carries allow="tools". The APIs also run only in origin-isolated documents, so a page’s origin stays fixed for a tool’s lifetime; a page that enables document.domain, for example with the Origin-Agent-Cluster: ?0 header, has them disabled.¹

For high-stakes or non-reversible actions, Chrome’s guidance is to set consequentialHint: true on the tool “so the agent or browser can request user confirmation before execution”, and the spec draft adds requestUserInteraction() to ask the user for input while a tool runs.¹⁹ Sites implementing transactional tools (payments, bookings, form submissions with consequences) should set the hint. Tools run visibly on the page rather than in the background,¹ and OpenAI says a site tool cannot authorise ChatGPT to take sensitive actions on a user’s behalf.⁷ Shopify’s checkout tools apply the same rule from the site side: agents must show the buyer the order and total and “get their permission to place it”, and Shopify states that neither an agent’s Web Bot Auth signature nor the checkout reaching its ready-to-complete status counts as that permission.¹¹

Prompt injection and data exfiltration risks are noted in the W3C spec as unresolved: it says it “cannot define precise mitigation strategies that agents or user agents must provide”. Its view of exposure is that WebMCP “does not inherently expand the attack surface”, because the functionality usually already exists in the site’s interface, though a tool may run different validation from the interface path.²⁰ Chrome’s guidance for a tool that returns user-generated or externally sourced content is to add untrustedContentHint, which labels the output as untrusted so the agent treats it with more scrutiny.¹⁹ During the experimental phase, avoid exposing sensitive internal data through WebMCP tools.

How does Lighthouse audit WebMCP?

Chrome’s Lighthouse 13.3.0 (7 May 2026) added an Agentic Browsing category and moved it out of experimental into the default config, so it now runs automatically in Lighthouse on Chrome 150 or later, and in PageSpeed Insights within roughly two weeks of each Lighthouse release.²¹ It is the first mainstream diagnostic that scores a page for agent readiness rather than human experience, and three of its audits target WebMCP directly:

• WebMCP form coverage: flags forms that could carry WebMCP annotations so agents can identify and submit them reliably.
• WebMCP tools registered: lists the tools registered via document.modelContext at the time of analysis.
• WebMCP schemas are valid: surfaces errors or warnings in tool schemas, since agents cannot call a tool whose schema they cannot parse.

The category reports a ratio of passed checks (for example 3/3) rather than a weighted score out of 100, because the standards for the agentic web are still emerging and the stated intent is to provide actionable signals rather than a ranking.²¹ On a site with no WebMCP implementation, the three WebMCP audits show as “not applicable” or unscored rather than failing. They only run at all on Chrome 150 or later with the WebMCP origin trial enabled, and Chrome weights them low while the standard stabilises.²¹

Two points of caution. The audit is a Chrome developer diagnostic, not a Google Search ranking signal: Chrome’s scoring page says the intent is “actionable signals rather than a definitive ranking”,²¹ and Google’s AI optimisation guide does not mention Lighthouse.² And the alongside checks in the same category (a well-formed accessibility tree, low Cumulative Layout Shift, and llms.txt presence) reflect general agent navigation readiness, not WebMCP specifically. Treat the WebMCP audits as a build-time correctness check for an implementation you have already decided to make, not as a reason to implement WebMCP.

What to do now

Informational and editorial sites: no implementation needed. Monitor adoption as the spec matures. The primary investment remains content quality, with structured data where it earns rich results.

E-commerce, travel, booking, and service sites: worth prototyping now. Review the Chrome for Developers WebMCP documentation and identify two or three high-value actions that would most benefit from agent invocation. ChatGPT’s desktop browser has called WebMCP tools since August 2026,⁷ and the Chrome 149 origin trial (from June 2026) lets sites run the API on live traffic. Prototyping now means being ready as more agents follow, not expecting immediate traffic impact. Shopify merchants are the exception: the platform already registers catalogue, cart and checkout tools on every Liquid storefront,¹⁰ so there is nothing to build for those actions. What the merchant controls is the checkout type, which decides whether an agent can finish the order or has to hand the buyer back to the page.¹¹

All sites: ensure existing structured data is solid. It keeps its documented jobs: rich-result eligibility in Google Search,² and, as structured content, helping Microsoft’s AI answers interpret a page.⁵ WebMCP supplements it; it does not replace it, and for most sites schema is the more established of the two investments.

Frequently asked questions

Does WebMCP affect search rankings?

Neither Chrome’s WebMCP documentation nor Google’s AI optimisation guide connects WebMCP to search rankings or AI citations.¹² WebMCP operates at the action layer, not the indexing or retrieval layer.

Is WebMCP the same as MCP?

Related but distinct. MCP (Model Context Protocol) is the broader Anthropic-developed standard for connecting AI models to external tools and data sources. WebMCP is a browser-native adaptation of that protocol, letting websites expose tools directly to in-browser AI agents via the document.modelContext API (renamed from navigator.modelContext, which was deprecated in Chrome 150). WebMCP uses MCP’s tool definition format but operates in a browser context rather than a server-to-server one.²²

Do I need WebMCP to appear in AI search answers?

No. Appearing in AI-generated answers is a retrieval and content quality question, covered by agentic search and GEO principles. WebMCP is about agent actions, not agent citations.

Will Google’s agentic booking feature use WebMCP?

Google’s agentic booking (placing calls to businesses on a user’s behalf) uses Google’s own infrastructure and is not directly tied to WebMCP. The two are complementary in intent but separate in implementation. WebMCP provides a standardised open protocol; Google’s agentic booking is a proprietary feature that may or may not adopt WebMCP as the standard matures.

Footnotes

1. WebMCP — Chrome for Developers ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹²

2. Optimizing your website for generative AI features on Google Search — Google Search Central. “Structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add.” The guide’s agentic section names UCP and browser agents; it does not mention WebMCP or Lighthouse. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

3. Introducing the Model Context Protocol — Anthropic ↩

4. WebMCP is available for early preview — Chrome for Developers ↩

5. How AI Search Is Changing the Way Conversions are Measured — Bing Webmaster Blog, Fabrice Canel and Krishna Madhavan, 20 November 2025: “Structured content, like schema-marked product pages, FAQs, and comparison tables, helps AI systems interpret and summarize your content more effectively.” ↩ ↩²

6. Google Search’s I/O 2026 updates: AI agents and more — Google ↩

7. Using site tools in the ChatGPT desktop app — OpenAI Help Center. Undated; Search Engine Journal reported the launch on 27 August 2026. ↩ ↩² ↩³ ↩⁴

8. Evolving Atlas into ChatGPT for browser-based agentic work — OpenAI Help Center ↩

9. 15 updates from Google I/O 2026: Powering the agentic web with new capabilities, tools, and features in Chrome — Chrome for Developers, 19 May 2026. ↩

10. WebMCP tools — Shopify Developers. Launched 5 August 2026 per Shopify’s developer changelog. ↩ ↩² ↩³

11. Checkout WebMCP — Shopify Developers. Launched 28 September 2026 per Shopify’s developer changelog. ↩ ↩² ↩³

12. The next evolution of the Gemini app — Google ↩ ↩²

13. Linux Foundation launches the Agent2Agent Protocol Project — Linux Foundation ↩

14. A2A Protocol Surpasses 150 Organizations, Lands in Major Cloud Platforms — Linux Foundation ↩

15. Read Sundar Pichai’s remarks at the 2026 National Retail Federation — Google Blog ↩

16. Under the Hood: Universal Commerce Protocol — Google Developers Blog ↩

17. Agentic Resource Discovery: let agents search — Hugging Face ↩

18. Agentic Resource Discovery Specification v0.91, agenticresourcediscovery.org, section 5.1 on the well-known URI and its predecessor path. ↩

19. WebMCP tool security — Chrome for Developers ↩ ↩²

20. WebMCP, Draft Community Group Report — W3C Web Machine Learning Community Group, section 6, Security and Privacy Considerations. ↩

21. Lighthouse agentic browsing scoring — Chrome for Developers ↩ ↩² ↩³ ↩⁴

22. When to use WebMCP and MCP — Chrome for Developers ↩