WebMCP (experimental)

WebMCP lets a browser agent call Varendo's registered tools on the page. The agent proposes an intent; the same application validates it and renders the result. The public Playground needs no API key.

1. Enable the browser experiment#

Varendo's native journey was tested with Chromium 153.0.8010.12 and the WebMCPTesting flag. Browser support is experimental; a version number alone does not mean the API is enabled.

1. Open chrome://flags/#enable-webmcp-testing in Chrome.
2. Set WebMCP for testing to Enabled, then relaunch.
3. Open the Playground and choose Browser agent.

The setup status should say Browser API available. If it says unavailable, check the flag, browser policy, and the document's permissions. Manual tasks remain available in every supported browser.

Production sites can instead enroll for a domain-specific origin-trial token. Enrollment and deployment are owned by the site maintainer; this guide does not imply that an origin trial is enabled on every Varendo domain. See Chrome's setup instructions.

2. Register and call the tools#

1. Choose Enable WebMCP. The status should report registered 3 tools.
2. Copy the example prompt and give it to your browser agent. It must be able to discover and invoke WebMCP tools in the current tab. Enabling the flag alone does not install an agent.
3. Ask the agent to show Engineering employees. The result changes on the page; Inspect shows the accepted request, result, and chosen presentation.
4. Try the workspace and entire-page demos. The connection stays active while the same endpoint targets the current registered demo.
5. Choose Disconnect or Reset playground to unregister the tools and cancel work owned by that connection.

The website does not call a model. Your browser agent owns its model, credentials, and any model charges. For a desktop MCP client or server-side BYOK, run the local example.

Detect support in your app#

Call detectWebMcp with the real document. It reports support when the document exposes modelContext.registerTool. The adapter never reads globals itself; you pass the document in. Mark the evidence native only when the browser supplied the capability.

capability.ts

import { detectWebMcp, registerWebMcpTools } from '@varendo/agent/webmcp';

const hostDocument = typeof document === 'undefined' ? undefined : document;
const capability =
  hostDocument === undefined ? detectWebMcp() : detectWebMcp({ document: hostDocument, evidence: 'native' });
if (capability.supported) {
  const registration = await registerWebMcpTools({ endpoint, document: hostDocument, evidence: 'native' });
  // Retain registration.value.adapter and close it with the session.
} else {
  showManualControls();
}

Keep a manual fallback#

The same resource must remain fully usable through buttons, filters, forms, and routes. Native registration is an optional communication surface, not a dependency for rendering or authorization.

Read the evidence correctly#

Chrome's imperative API guidance defines tool registration, discovery, execution cancellation, and disposal. As of Chrome 153, unregistering a tool does not cancel work already in flight, so Varendo also checks the endpoint lifecycle after execution and rejects late results. Cross-origin tools additionally require the browser tools Permissions Policy and explicit origin exposure.