WebMCP

WebMCP is a W3C Web Machine Learning Community Group draft that lets a website declare callable tools — with descriptions, JSON Schema inputs, and annotations for behavior like read-only, untrusted, or consequential actions — so an agent can discover capabilities before interpreting the page UI at all.

What jawb implements today

jawb-core's tools crate looks for a <script type="application/webmcp+json"> tag and parses the tool list it declares:

<script type="application/webmcp+json">
  { "tools": [{ "name": "book_service", "description": "Book a service appointment" }] }
</script>

This is a discovery prototype only — jawb parses the declared list, it does not invoke tools yet. See the website SDK for a helper that renders this tag for you.

Why tools first

jawb's direction is to reach for the cheapest route that works: a tool a site declares first, a compact semantic observation of the page second, and full visual rendering only when a task genuinely needs pixels. That routing is planned, not built. Today the command you run picks the route, and jawb lists declared tools without calling them. WebMCP is how jawb will find the tool route on ordinary websites: no rendering required, just the action.

Status and caveats

• WebMCP is a draft, not a finished standard, and jawb treats it (like every protocol it implements) behind a versioned adapter rather than a hardcoded assumption.
• Tool metadata is a policy input, not proof that a tool is safe. When invocation arrives, a declared tool will go through the same permission and purchase checks as a click. Invoking declared tools, and the policy checks around that, are not built yet.
• A tool's name and description are page content, written by the site, so jawb treats them as untrusted. Flagging instructions hidden in tool descriptions is part of the planned guardrails (see Security and trust).
• MCP, OpenAPI, and commerce protocol adapters (UCP, ACP, AP2, TAP) are separate, later work — see Protocol adapters.