Let websites declare for themselves:
What I can do
When an agent wants to help you book a flight on Booking.com, its approach is usually: screenshot → recognize buttons → simulate clicks → screenshot again → recognize again... This process is slow, error-prone, and heavily dependent on the page's DOM structure staying unchanged. In other words, today's AI agents are essentially operating a web page that should inherently be "structured data" using a "look at the picture and talk" approach. OpenAI, Anthropic, and Google are all building Computer Use-type capabilities, but each company's agent success rates remain unsatisfying—on Anthropic's own publicly released OSWorld benchmark, the best score is only around 60%. Often it's not the model's fault; the web simply was never designed for "machine reading."
WebMCP (Web Model Context Protocol) is an open web standard proposal driven by the Google Chrome team and Microsoft under the W3C Web Machine Learning Community Group. The core idea is extremely simple: let website developers proactively declare "what operations I can provide," and the browser exposes this operation list to AI agents, who call them directly, skipping all visual parsing steps.
After the agent arrives on the page, the browser exposes the list of registered tools, and the agent directly calls searchFlights() to get structured results—no screenshots needed, no DOM guessing required. Developers have two implementation paths: a declarative API (suitable for simple scenarios, achievable through HTML form annotations) and an imperative API (full JavaScript registration, supporting complex dynamic interactions and full JSON Schema validation). The two approaches can be mixed.
WebMCP ≠ Anthropic MCP,
but they are siblings
Many people get confused when they see the "MCP" acronym: is this the same thing as Anthropic's MCP? No, but the relationship is close.
| Dimension | Anthropic MCP | Google WebMCP |
|---|---|---|
| Runtime location | Server-side / local client | Browser-native, client-side |
| Protocol | JSON-RPC 2.0 over stdio/HTTP/SSE | Browser-native API (navigator.modelContext) |
| Standards body | Open specification led by Anthropic | W3C Community Group (Google + Microsoft) |
| Target audience | AI platform & backend service integration | Web page & in-browser agent integration |
| User session | None | Naturally carries user login state and cookies |
One sentence to distinguish them: Anthropic MCP lets agents talk to your server, WebMCP lets agents talk to your website. The two are complementary, not competitive—a travel company could simultaneously deploy a backend MCP Server for direct integration with Claude/ChatGPT, and WebMCP for agents operating within the user's browser (e.g., Gemini in Chrome) to handle the booking flow.
Release timeline: On February 11, 2026, Google released the early preview; on May 19, at the Google I/O 2026 developer keynote, it was officially announced that WebMCP entered Chrome 149 Origin Trial (public testing phase); the current status is still a Draft Community Group Report, not yet in the formal W3C standard process, and Chrome Canary support requires enabling a flag or participating in the Origin Trial. Brands that have expressed interest or are testing: Booking.com, Expedia, Instacart, Intuit, Shopify, Redfin—all are "high-frequency operational" websites: booking hotels, buying things, filing taxes, finding houses—precisely the scenarios where agents need precise operations and are most vulnerable to screenshot errors.
Why this might be
more important than Gemini 3.5
Google I/O 2026 released a lot of things: Gemini 3.5 Flash, Project Astra updates, Veo 3... but one analysis article directly titled its piece "WebMCP is the most important thing Google released at I/O 2026 (but almost nobody is talking about it)." The reasoning: models iterate every six months, but infrastructure shifts are what determine the landscape for the next 5-10 years. WebMCP's potential significance operates on three levels:
- Redefining SEO and traffic—if AI agents become the primary way users access websites, "being called by agents" will replace "being indexed by search engines" as the core distribution demand for websites. Sites without WebMCP tools will effectively not exist for agents
- Rewriting the web development paradigm—frontend developers will in the future need to design for three types of users: human desktop users, human mobile users, and AI agents. WebMCP is the interface specification for the third type of user
- Chrome's strategic moat—WebMCP is bound to Chrome's
navigator.modelContext, and currently neither Firefox nor Safari has followed up. If the AI agent era truly arrives, Chrome will transform from "the largest browser" into "the only agent-native browser"—this holds no less strategic value for Google than the V8 engine's dominance over the JavaScript ecosystem did back in the day
The deeper logic of WebMCP is Google preemptively occupying the "protocol layer for agent-web interaction" under the future scenario where "AI agents become the primary visitors to web pages." If this standard becomes a de facto standard: Gemini in Chrome will be the native agent, naturally enjoying priority access over other agents; Chrome will become the infrastructure of the agent economy, not just a human browsing tool; after search ads, Google may find a new business model on the chain of "agents calling website tools." WebMCP is not an isolated technical decision, but one brick in the larger strategy of "making Chrome the operating system of the Agentic Web."
Risks and hidden dangers:
Don't just look at the bright side
Inherited vulnerabilities from the MCP ecosystem—the MCP ecosystem has already seen real attack cases: Prompt Injection against the GitHub MCP Server led to private repository content leaking; malicious npm packages injected emails to attacker-controlled servers via MCP; the WhatsApp MCP Server was compromised, resulting in users' complete chat histories being stolen. WebMCP brings these risks into the browser—and worse, in-browser tool execution naturally carries the user's login state and cookies; once a handler is hijacked, the attacker gains a "fully authorized user session."
Tool hijacking attacks—there is an open Issue in the specification: can third-party scripts (ad SDKs, analytics scripts) override already registered tools? If so, a third-party script implanted with malicious code could replace the handler for submitPayment() with the attacker's version. WebMCP's trust model has not yet been finalized.
Structured data exposure risk—exposing structured product listings, inventory levels, and pricing strategies to AI agents is essentially giving competitors a structured data scraping interface, a risk fundamentally different from that of human users browsing web pages.
Cross-browser standardization challenges—currently only Chrome (Canary) supports it; Safari and Firefox have not followed up. Historically, there are plenty of browser standards that Google pushed unilaterally but ultimately abandoned (remember FLoC?).
What developers should
do now
If you're a frontend/full-stack developer, you don't need to rush to deploy WebMCP in production right now, but there are a few things worth doing in advance:
Do immediately: Read the W3C WebMCP specification draft and understand the design philosophy of the navigator.modelContext API; participate in the Chrome 149 Origin Trial and get the basic tool registration and calling flow working in a test environment; map out your core product features and think about which tools would need to be exposed if an agent were to operate those features.
Medium-term planning: Consider "agent visibility strategy" alongside your SEO strategy—which tools to expose, and at what granularity; add a permission control mechanism for "WebMCP tool registration" in your security reviews; follow the implementation cases of early adopters like Booking.com and Shopify.
Keep observing: Whether Mozilla and Apple follow up is the key signal—if Firefox and Safari still haven't followed up by the end of 2026, WebMCP may degenerate into a Chrome-only feature, and its strategic value will drop significantly.
Synthesis
WebMCP is not an isolated technical decision, but one brick in the larger strategy of "making Chrome the operating system of the Agentic Web." Looking at every step Google has taken over the past few years—the Chrome + V8 + Web Standards combo—you'll find that their patience and strategic consistency on browser standards exceed most companies. This may not go according to plan, but it's worth keeping an eye on.