WebMCP

WebMCP is an experimental browser API that allows a web application to register JavaScript tools for compatible AI agents. In the current draft, a page uses document.modelContext.registerTool() to describe an action, its input schema, and the code that executes it.

That definition is deliberately narrow. WebMCP is not a generic label for every “MCP-compatible website,” not a metadata file, and not a search-ranking feature. The current specification is a Draft Community Group Report, not a W3C Standard or a standards-track recommendation.

What The Current Draft Defines

The imperative API attaches a ModelContext to the document. A page can register a named tool while it is active:

await document.modelContext.registerTool({
  name: "filter-products",
  description: "Filter the visible catalog by category and maximum price.",
  inputSchema: {
    type: "object",
    properties: {
      category: { type: "string" },
      maxPrice: { type: "number" }
    },
    required: ["category"]
  },
  annotations: { readOnlyHint: true },
  execute: async ({ category, maxPrice }) => {
    return applyVisibleFilters({ category, maxPrice });
  }
});

The registered action can reuse page state and existing client-side logic. The draft also defines origin exposure options, lifecycle behavior, and annotations such as readOnlyHint and untrustedContentHint.

The declarative form-based part of the specification is still unfinished. Do not describe automatic form conversion as a stable production capability.

WebMCP Versus MCP And OpenAPI

Surface	Runs where	Describes	Best fit
WebMCP	Open browser document	Page-context JavaScript actions	Visible, user-supervised UI tasks
MCP server	Server or local runtime	Tools, resources, and prompts	Agent integrations independent of one open page
OpenAPI	HTTP service documentation	Endpoints, operations, schemas, and security schemes	Public or authorized API clients
Ordinary HTML	Browser and search crawler	Human-readable content and controls	Discovery, explanation, and fallback UI

These layers can coexist. A product catalog may document an HTTP API with OpenAPI, expose selected workflows through an MCP server, and experiment with a WebMCP filter tool on the visible catalog page. One does not replace the others.

Appropriate First Uses

Good early experiments are narrow, visible, reversible, and already supported by the normal UI:

• Filter or sort items that are already on screen.
• Populate a draft form and let the user review it.
• Run read-only analysis against data visible to the signed-in user.
• Navigate a complex interface through existing checked actions.

Avoid starting with payment, account deletion, publication, permission changes, or any action whose authorization only exists in the tool description.

Security Boundaries

Tool metadata becomes part of an agent’s context, so descriptions and outputs can carry untrusted instructions. A useful implementation review covers:

1. Authorization: the server must enforce the same permissions as the UI.
2. Side effects: irreversible or paid actions require explicit confirmation.
3. Input validation: accept only fields and values the underlying action supports.
4. Origin exposure: expose tools only to intended origins.
5. Untrusted content: label and sanitize tool output that contains user-controlled text.
6. Fallback: keep the normal accessible UI because client support is not guaranteed.

An agent-readable description is documentation, not a security boundary.

SEO Impact

There is no documented WebMCP ranking signal, structured-data feature, or AI-citation benefit. A page still needs crawlable text, stable URLs, internal links, useful content, and a good page experience. Build WebMCP only when it improves a real user task for a supported agent runtime.

For a deeper implementation and product review, read WebMCP for SEO and Product Websites. Compare it with Model Context Protocol, API Catalog SEO, and A2A Agent Card.

Validation Checklist

• Confirm the target browser or agent actually implements the draft.
• Test registration, unregistration, cancellation, and duplicate names.
• Verify authorization on the underlying server request.
• Test malicious text in tool descriptions, inputs, and outputs.
• Confirm high-impact actions cannot run without user review.
• Preserve an accessible non-agent workflow.

Primary Sources

• WebMCP Draft Community Group Report
• WebMCP proposal and explainer
• WebMCP source repository