What Is WebMCP? Preparing Websites for AI Agents

Chrome updated its WebMCP guidance on October 7, 2026, and the community draft is dated October 27. These primary-source developments informed SoSho’s editorial choice. They establish active work on the topic, rather than a ranking of search demand or a promise of commercial results.

A useful business question is what a customer’s assistant should be able to do on your site. Finding an appropriate service or filtering products may be enough for an initial experiment. Following the trend does not require replacing your entire website or visual identity.

Chrome describes declaring a capability’s purpose and inputs, instead of relying entirely on an agent interpreting a sequence of clicks. The proposal includes JavaScript tools and HTML-form annotations. Treat this as progressive enhancement: people still need a usable form and visible controls.

Consider a hypothetical workspace-booking site. A customer requests a six-person room for Tuesday. A search tool returns available options; the customer then reviews terms and confirms a choice. This is SoSho’s proposed pilot scenario, not a claim about a completed client project.

The community draft concerns tools exposed by a web application in its page context. It explicitly is not a completed W3C standard. An experimental browser integration should therefore not be sold as a universal replacement for a server API.

Our editorial assessment separates agent usability from search discovery. Clear service descriptions, readable HTML and sensible internal links remain independent work. Merely registering a tool does not justify promising better search rankings or more citations in AI answers.

Before choosing a library, write one scenario and an acceptance criterion. For example: given a service category and city, the agent should return three relevant options. Compare the ordinary interface with the structured route. Look for ambiguous inputs and incorrect outcomes rather than maximising the number of registered tools.

For a small team, the pilot can end with a short report covering successful requests, missing information and whether customers could correct the result. If unclear copy or a broken form is the underlying problem, address that obstacle first.

The community draft discusses prompt injection, privacy leakage and divergence from user intent. Our implementation recommendation is to retain server validation and account-based permissions rather than rely on a persuasive tool description.

In the booking example, displaying a price and committing a reservation are different operations. Show the selection, cost and terms before the consequential action. Use synthetic records and test accounts in the pilot so cancellation and incorrect inputs can be examined without a real booking.

Current Playwright documentation describes discovering and calling WebMCP tools with experimental browser support enabled. Tool descriptions and results are themselves untrusted input. Combine these checks with ordinary human-interface tests and record the tool and browser versions used.

Expand the pilot one step if it solves a real problem. Keep it as an experiment if it only makes an attractive demo. For a useful technical estimate, first list the customer tasks, their permission requirements and the human confirmation steps.