Browser Control with OpenClaw

OpenClaw can work on web pages next to the chat. v2026.8.2 added browser control that continues when the Gateway is not running. v2026.9.3 added live browser automation. This is useful and easy to expose.

Official security notes: require auth for loopback browser-control HTTP routes, and auto-generate gateway.auth.token when browser control starts without auth. Treat page content as untrusted.

What “Gateway optional” means

The Gateway is still the usual control plane for channels and agents. Browser control in 8.2 can keep a session attached to a local browser even if you stopped the Gateway process. That is for desktop-style work, not an excuse to bind browser-control HTTP to 0.0.0.0.

Safe setup

  1. Keep Gateway and browser-control ports on loopback. Remote access via SSH tunnel or Tailscale. Network isolation.
  2. Confirm a token exists before you start browser control. If the product generates one, store it like any other secret. Credentials.
  3. Do not feed raw HTML from the page back into a high-privilege agent without a wrapper. Prompt injection from the live web is the default threat. Prompt injection.
  4. Pair with workspace-only limits so media reads cannot leave the folder you intended. October patch.

Related