What Is WebMCP?

What is WebMCP?

WebMCP is a browser-facing way for a website to describe useful capabilities to an AI agent in the context of the page. It complements ordinary HTML, forms and application APIs. A visitor can continue using the human interface, while an agent can discover a narrow action with a name, description, structured input and result. WebMCP is not a new permission system, an SEO shortcut or a replacement for server-side application logic.

How a WebMCP tool works

A good tool starts from a real user goal, such as finding a product, locating support information, preparing a quote or checking appointment availability. The page exposes a contract; the agent chooses it, provides input and receives a result. The website remains responsible for input validation, authentication, authorization, consent, rate limits, error handling and the final state change. The tool should drive the same trusted workflow that a person can use.

The current API surface

The current documented imperative path uses document.modelContext and registerTool(), with a tool name, description, input schema and execution function. A minimal design might register search_products with a required query string and a read-only result. Feature-detect the API in browser code, keep registration out of server-only rendering, and preserve a usable fallback when the browser does not expose the surface. navigator.modelContext should be treated only as a legacy compatibility signal.

What makes a useful first tool?

Start narrow. “Find a black laptop under €900” is a better first task than “operate the shop” because the inputs and result can be tested. “Show my invoices from March” can be read-only but still requires account authorization. “Prepare a quote for 500 units” can collect structured information for human review without submitting an order. MFA reset, booking, checkout, payment and order creation can affect identity, money or external commitments, so the commit must remain confirmation-required.

What WebMCP does not solve

A tool declaration does not make untrusted input safe, grant access to private data or guarantee that an endpoint behaves correctly. Do not put secrets in browser code. Do not assume that a hidden field, disabled button or client-side route protects a business action. Validate at the handler and again on the server; check authorization for the current user; make consequential effects visible; and log or audit the operation according to the application’s requirements.

WebMCP and search visibility

Server-rendered explanations, useful headings, accessible forms, descriptive links and stable canonical URLs make this topic understandable to people and search engines. They do not register a tool by themselves. Optional discovery conventions such as llms.txt or schema.org actions can provide context, but they are not official WebMCP requirements. Technical SEO should support the product workflow rather than promise agent access that has not been tested.

How to check and implement it

Use the Checker for static source evidence, the Simulator for mapping a natural-language task to an action, and Autopilot for prioritisation. The Generator creates a reviewable schema and code scaffold; the Validator separates static findings from runtime verification. For a production workflow, the Implementation Service can review the full application boundary.

FAQ: is WebMCP stable everywhere?

No blanket browser claim should be made. Browser channel, page lifecycle, permissions, schema and application authorization all matter. A static reference is evidence to investigate, not proof that an agent can register, invoke or complete a task. Review the browser-support reference and label results as detected, suggested, runtime verified or not run.

A simple mental model

Think of WebMCP as a typed bridge between a user goal and an existing web workflow. The bridge describes intent; it does not own the business truth. Search results still come from the catalogue, invoice access still comes from the account system, and an order is still governed by the commerce backend. This model prevents a page from exposing a decorative control that cannot be trusted in production.

A sensible rollout

Choose one page, one action and one owner. Define the schema, result, authorization and confirmation point; test the ordinary form; add the current browser registration; then compare agent and human outcomes. Record what is detected, suggested, runtime verified and not run. Continue with the Simulator, Autopilot, Generator, Validator and Implementation Service as the workflow matures.