Files
openclaw/docs/tools/chrome-extension.md
Peter Steinberger 4efcea1fd2 feat(browser): send pages to OpenClaw from the Chrome extension (#111158)
* feat(browser): send pages to the main session from the Chrome extension

One-click page share in the OpenClaw Chrome extension: toolbar popup with an
optional note, page/selection context menu, and Alt+Shift+S. Capture is
selection-first with a readability heuristic, X/Twitter thread extraction, and
Google Docs plain-text export via the user's session cookies. Payloads ride the
existing paired relay WebSocket as a new pageShare message; the gateway-only
page-share sink wraps page text in the external-content safety boundary, then
enqueues a main-session system event and requests an immediate heartbeat
(hooks/wake semantics). Node-hosted relays report a clear unsupported error.

Capture heuristics adapted from Nat Eliason's MIT-licensed send-to-openclaw.

Co-authored-by: Codex <codex@openai.com>

* fix(browser): keep page-controlled metadata inside the share safety boundary

Review findings: move title/URL inside wrapExternalContent (a hostile <title>
must not become trusted header text), prefer the user's selection over the
full Google Docs export, and pass the context-menu selectionText through so
iframe selections and selections cleared during relay reconnect still win.

* fix(browser): bind context-menu shares to the click-time document

Selection shares from the context menu now send the click snapshot directly
(no recapture), so navigations during relay reconnect cannot mislabel the
source and iframe selections are preserved. The Google Docs selection probe
scans all accessible frames before falling back to the full-document export.

* test(browser): expect the page-share handler in relay server args

* fix(browser): probe only the main frame for Google Docs selections

All-frame injection rejects wholesale when one frame is inaccessible and
returns child frames in nondeterministic order, so the probe now reads the
main frame only. Child-frame selections still share correctly through the
context menu's click-time selectionText; toolbar/shortcut entry sends the
full page for that case (named tradeoff in the code comment).

* fix(browser): satisfy page-share CI gates

---------

Co-authored-by: Codex <codex@openai.com>
2026-07-19 01:07:28 -07:00

10 KiB

summary, read_when, title
summary read_when title
Chrome extension: let OpenClaw drive your signed-in Chrome with no remote-debugging prompt
You want an agent to drive your real signed-in Chrome from your phone
You keep hitting the Chrome "Allow remote debugging?" prompt with nobody at the desk
You want to understand the security model of browser takeover via the extension
Chrome Extension

Chrome extension

The OpenClaw Chrome extension lets an agent control your signed-in Chrome tabs without launching a separate managed browser, and without Chrome's blocking "Allow remote debugging?" prompt.

This matters when you drive OpenClaw from a phone (Telegram, WhatsApp, etc.): the user profile attaches over Chrome's remote-debugging port, which pops a desktop consent dialog nobody can click when you are away. The extension uses the chrome.debugger API instead, so the only in-page hint is Chrome's dismissible "OpenClaw started debugging this browser" banner.

This is the same shape used by Anthropic's Claude in Chrome and OpenAI's Codex Chrome extensions.

How it works

Three parts:

  • Browser control service (Gateway or node host): the API the browser tool calls.
  • Extension relay (loopback WebSocket): a small server the control service starts on 127.0.0.1. It presents a Chrome DevTools Protocol endpoint to OpenClaw and speaks to the extension. Both sides authenticate with a host-local token (see below).
  • OpenClaw Chrome extension (MV3): attaches to tabs with chrome.debugger, forwards CDP traffic, and manages the OpenClaw tab group.

OpenClaw only sees and controls tabs that are in the OpenClaw tab group. The group is the consent boundary: drag a tab in to share it, drag it out (or click the toolbar button) to revoke access instantly.

Install and pair

  1. Print the unpacked extension path:

    openclaw browser extension path
    
  2. Open chrome://extensions, enable Developer mode, click Load unpacked, and select the printed directory.

  3. Print the pairing string:

    openclaw browser extension pair
    
  4. Click the OpenClaw toolbar icon and paste the pairing string into the popup. The badge turns ON when the extension connects to the relay.

The pairing token is a host-local secret created on first use and stored under credentials/ in the state directory (mode 0600). Each machine that runs a browser — the Gateway host and every browser node host — owns its own token, so no credential has to travel between machines. To rotate it, delete the browser-extension-relay.secret file and pair again.

Use it

Select the built-in chrome profile in a browser tool call, or make it the default:

openclaw config set browser.defaultProfile chrome
{
  browser: {
    profiles: {
      chrome: { driver: "extension", color: "#FF4500" },
    },
  },
}
  • Share a tab: click the OpenClaw toolbar button on that tab (it joins the OpenClaw tab group), or drag any tab into the group.
  • The agent can also open new tabs; those land in the group automatically.
  • Revoke: click the button again, drag the tab out of the group, or dismiss Chrome's debugging banner. The agent loses access to that tab immediately.

Tab copilot side panel

After pairing the extension, click Open tab copilot in its toolbar popup. OpenClaw configures sidepanel.html for that exact Chrome tab; the manifest has no global side-panel path. Each tab therefore gets a separate panel document, Gateway session, message subscription, and typed browser-tool binding.

The panel does not place the page URL, title, DOM, or visible text in your message. It sends only the text you type. Browser actions carry a separate Gateway-authenticated binding containing the Chrome tab and CDP target, and the browser tool rejects attempts to replace that target or use browser-wide actions. Replies stay in the panel (deliver: false); they do not inherit a Telegram, Discord, or other channel route.

The copilot is a dedicated paired Gateway device with operator.read and operator.write scopes. On first use, inspect and approve its request:

openclaw devices list
openclaw devices approve <requestId>

The extension retains that device identity and the Gateway-issued device token, scoped to the canonical Gateway endpoint that issued them. Pairing a different Gateway creates separate identity, token, and session custody; credentials and sessions are never reused across endpoints. The extension does not persist the Gateway shared secret. A panel can subscribe only to its own tab sessions, and the Gateway filters those events before delivery.

If the Gateway connection drops during a run, the extension keeps durable custody of that run ID. On reconnect it aborts the unresolved run before re-enabling any panel, then reloads transcript history. This fail-closed step prevents browser actions from continuing unseen across a delivery gap.

Closing a tab immediately removes its live subscription, aborts any visible run, and marks that tab's session archived. If the Gateway is temporarily offline, the extension persists the pending archive and retries only when that same Gateway endpoint reconnects; it never sends an archive request to a different Gateway. After a browser crash, the next launch archives sessions left by the previous browser instance. Archived sessions reject new work, while their transcripts remain available in session history. Browser-copilot keys are thread sessions, so normal age and entry-count maintenance preserves them. The per-agent session disk budget still applies (default 2gb) and may evict the oldest sessions under pressure; see session maintenance.

The side panel currently requires either a Gateway-hosted extension relay or a direct remote Gateway relay. A loopback relay on a browser node cannot yet provide the node route required by the typed tab binding, so the panel denies that topology instead of falling back to browser-wide routing.

Send a page to OpenClaw

Use Send page to OpenClaw in the toolbar popup to share readable page text with your main OpenClaw session. You can add an optional note, use the page or selection right-click menu, or press Alt+Shift+S. OpenClaw prefers your current selection when one exists, enqueues the share as a system event, and wakes the main session immediately.

The tab does not need to be in the OpenClaw tab group. This is a one-shot, explicit share: nothing else on the page is exposed, and it grants no ongoing access. Google Docs are exported as plain text with your signed-in browser session, without Google API setup. X and Twitter threads are extracted without the surrounding interface chrome.

Page text is wrapped in OpenClaw's external-content safety boundary. Your optional note stays outside that boundary as your own instruction. Page text and selections are capped at about 120,000 characters and include a truncation marker when shortened.

Page sharing works when the extension relay is hosted by the Gateway, using same-host pairing or direct wss:// Gateway pairing. Node-hosted relays return a clear error for now. To remap the keyboard shortcut, open chrome://extensions/shortcuts.

Remote / cross-machine

Chrome does not have to run on the Gateway host. Three topologies work:

  • Same host (Gateway + Chrome on one machine): pair on that machine with openclaw browser extension pair. The relay is loopback-only. If the local Gateway uses TLS, pass its certificate hostname explicitly with --gateway-url wss://gateway-host.example; pairing never substitutes a loopback IP.
  • Direct to a remote Gateway (Chrome on your laptop, Gateway on a VPS, and nothing else on the laptop): on the Gateway, run openclaw browser extension pair --gateway-url wss://your-gateway.example.com. It prints a wss://…/browser/extension#<secret> string; load and pair the extension on the laptop. The extension connects straight to the Gateway over wss:// — no OpenClaw install, Node, CLI, or open inbound port on the laptop. This is the managed-hosting path.
  • Via a browser node host (Chrome on a machine already running an OpenClaw node): run pair on the node and pair locally; the Gateway proxies browser actions to the node over its existing authenticated node link.

The pairing secret is per host (the Gateway's, in the direct case), validated by the Gateway's /browser/extension route. For the direct path, serve the Gateway over TLS (wss://) so the pairing secret and CDP traffic are encrypted. The secret remains in the pairing string's URL fragment and is presented during the WebSocket handshake as a subprotocol credential, so normal proxy access logs do not receive it in the request URL. Ensure any reverse proxy preserves the standard Sec-WebSocket-Protocol header.

Diagnostics

openclaw browser status --browser-profile chrome
openclaw browser doctor --browser-profile chrome

doctor reports the Chrome extension relay check as failing until the extension popup shows Connected.

Security model

  • The relay binds loopback only; both WebSocket sides are authenticated with the derived token, and the extension side is origin-checked to chrome-extension://.
  • Direct Gateway pairing does not accept the relay token in the request URL; the bundled extension carries it in the WebSocket subprotocol list instead.
  • The agent can only see and drive tabs in the OpenClaw tab group. Your other tabs stay private.
  • Side-panel runs are scoped twice: Gateway delivery uses a per-session allowlist, and browser tools enforce the Chrome tab/target binding carried outside the prompt.
  • Compared with the user (Chrome MCP) profile, which exposes your whole signed-in browser once you approve the remote-debugging prompt, the extension keeps the shared surface scoped to a tab group you control at a glance.

See also: Browser for the full profile model and the managed openclaw and Chrome MCP user profiles.