[compnew2006]

What

This PR adds two tools, browser_webmcp_list and browser_webmcp_call. Some sites expose their own actions to agents through WebMCP, using document.modelContext.registerTool(...) or a <form toolname>. On those sites, an agent can now list the tools and call one by name with an input object, instead of clicking through the UI. Pages without WebMCP behave exactly as before.

How

• Transport. A self-contained page runner (extension/lib/webmcp-page.js) is evaluated in the page's main world over CDP.
  • The call is an awaited expression, because REPL mode does not unwrap a returned promise.
  • The runner returns a JSON string, and the handler (extension/handlers/webmcp.js) validates every field of it.
• Safety gates, checked against the live tool at call time:
  • consequentialHint tools and toolautosubmit forms need confirm: true.
  • A form that waits for the user's submit returns DECLARATIVE_NEEDS_USER; the agent shield would otherwise block that submit.
  • Declarative forms are found in the light DOM, in open and closed shadow roots, and in same-origin frames, via DOM.performSearch. If that scan fails, calls fail closed with FORM_SCAN_FAILED.
  • browser_shortcuts refuses to save confirm: true.
• Untrusted content. Tool names, descriptions, schemas and output are page-authored. The handler sets untrustedContent: true on them; the flag is never taken from the page.
• Timeouts and navigation. A timeout aborts the tool page-side, so it is never mistaken for a frozen tab. A navigation during a call returns a non-error status: "unknown".
• Frames. Tools with the same name in different frames are told apart by frame / frameUrl.

The design and its review history are in specs/webmcp-support/:

Stage	Verdict
Independent plan review	REJECT, then APPROVE WITH CHANGES
Independent audit	FAIL, then PASS WITH FINDINGS, then PASS

Verification

Check	Result
npm run build / typecheck / lint	exit 0
npm test	629 passed, 44+ of them new for WebMCP
npm run test:coverage	thresholds pass
git diff --check	clean
npm run smoke:webmcp	15/15

npm run smoke:webmcp runs the real handlers against an isolated headless Chrome 155 (temporary profile, --enable-features=WebMCP). It needs a local Chrome 149 or newer, so it is not part of npm test or CI.

Not done here

• An end-to-end run through a live daemon and extension in a daily-use Chrome. That needs chrome://flags/#enable-webmcp-testing, an extension reload, and npm run daemon:restart.
• No version bump.
• Closed-shadow-root detection relies on DOM.performSearch, an experimental CDP command. Re-run npm run smoke:webmcp after Chrome upgrades.

🤖 Generated with Claude Code

[qodo-code-review]

ⓘ Qodo reviews are paused because the subscription is no longer active. Ask your workspace admin to reactivate the subscription to resume reviews. Manage billing

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

• Configuration used: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: e3a99bd3-ff2a-4d2d-9dff-91c6e6130aa7

---

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

---

_{Comment @coderabbitai help to get the list of available commands.}