@clawnify/browser-tabs

Per-agent browser tab isolation for OpenClaw, when multiple agents share one browser.

By Clawnify.

The problem

Run several OpenClaw agents against a single shared Chrome (one CDP endpoint) and they share one flat tab namespace. When they work at the same time they step on each other's tabs — you'll see tab not found, action targetId must match request targetId, and stale refs, because one agent's snapshot/act lands on a tab another agent just changed.

This plugin scopes the browser per agent: an agent may only act on tabs it opened, while all agents keep the same shared browser profile (so shared logins/cookies are preserved).

What it does

Entirely at the tool-call boundary — no fork of the browser plugin, pure string/Map work (no child_process/eval), so it installs as a normal external plugin:

• before_tool_call — blocks a targetId/tabId/label that resolves to another agent's tab, and injects the agent's own last tab when the model omits a target on snapshot/act/screenshot/console.
• after_tool_call — records ownership when an agent opens a tab (released on close), keyed by the calling agent (ctx.agentId).

Ownership is keyed by the tool's stable tabId (raw DevTools targetIds change on navigation) and kept at module scope so it survives the host re-running the plugin's register().

Install

openclaw plugins install @clawnify/browser-tabs --pin

Then enable it in your OpenClaw config (plugins.allow / plugins.entries), with prompt-injection hooks allowed (it rewrites tool params):

{
  "plugins": {
    "allow": ["browser", "clawnify-browser-tabs"],
    "entries": {
      "clawnify-browser-tabs": {
        "enabled": true,
        "hooks": { "allowPromptInjection": true }
      }
    }
  }
}

The browser plugin stays enabled — this wraps it, it does not replace it.

Config

{
  "enabled": true, // master switch (default true)
  "debug": false   // log each observe/enforce decision (default false)
}

Scope

This delivers action isolation — agents can't act on each other's tabs and default onto their own. It does not hide foreign tabs from the action=tabs list (a plugin cannot rewrite the live tool result the model sees). Complete list-scoping belongs in the browser tool itself and is best solved upstream.

Ownership is in-memory: a full gateway restart resets it (it self-heals as agents reuse tabs).

License

MIT.

Built by Clawnify — managed OpenClaw AI agents for non-technical teams.