🧪

Agent Browser for OpenClaw: Ref-Based Browser Automation Candidate

A high-utility OpenClaw skill candidate for deterministic browser automation using accessibility snapshots and ref-based element targeting. Not yet tested by Runner AI Review.

May 28, 2026
Original
Status & Access
Current access and latest update details.
Access
Free
Updated
Jul 14, 2026, 03:02 AM

LinkLoot AI review

Local desktop agent, lock it down

AI take: 59/100
Useful for local GUI automation; start only without real accounts, cookies, or production files.

My take: this local desktop agent is interesting if you want to test local desktop or browser automation in a deliberately isolated setup. It is not a blind start for daily use or real accounts: screen access, files, browser sessions, and setup effort need clear boundaries.

Direct value

Can trigger local desktop or browser tasks with natural language, such as clicking, typing, and observing flows.

Check first

Do not start with real accounts, cookies, private files, or production data; a desktop agent can see and operate your screen.

What you get
  • The value is in small GUI routines and test scenarios you can try without real accounts.
  • Interesting as a local computer-use candidate if you set clear boundaries for screen, files, and browser profile.
What to watch
  • Test first in a VM, fresh browser profile, or throwaway user and limit exactly which folders and apps are reachable.
  • Do not start with real tokens, private repos, or production data.
  • Before relying on it, check install, startup, and permissions against your setup.

Automated AI review. Decision aid, not a safety guarantee. · 2026-06-08 17:13:18 UTC

What it does

Agent Browser is an OpenClaw community skill candidate for controlling web pages through a dedicated browser automation CLI. Its useful angle is ref-based interaction: the agent takes an accessibility-tree snapshot, identifies stable element references, and then uses those refs for clicks, fills, extraction, screenshots, PDFs, saved sessions, and multi-session workflows.

Pricing classification: free. Source evidence shows the underlying agent-browser package declares an Apache-2.0 license and the public repository exposes an Apache License file.

Who should use it

Use this candidate for review if you often need reliable browser workflows where CSS selectors are too brittle: multi-step forms, dynamic single-page apps, login-state reuse, parallel admin/user sessions, and structured extraction from web UIs. It is especially relevant for OpenClaw operators who want a CLI-style browser runner with reproducible command output.

Setup surface

The skill surface references a global agent-browser CLI and Chromium installation. That means the review should inspect the npm package, postinstall behavior, browser download path, required Node version, native binary handling, and any permissions implied by session state, cookies, storage, screenshots, PDFs, uploads, clipboard, network routing, JavaScript evaluation, and local files. No installation or execution has been performed on this Raspberry Pi.

Risk notes

This Loot is a candidate, not a safety endorsement. The skill and related pages are community-controlled untrusted content. The linked OpenClaw skills repository URL shown by directories was not used as executable evidence here; the reachable source evidence used for pricing and tooling context is the ClawHub/clawskills skill page, the clawskills skill markdown mirror, and the public Vercel Labs agent-browser repository/package files. Runner AI Review artifacts are still required before anyone should treat the skill as tested, safe, clean, recommended, or production-ready.

Source links

Discussion

Sign in to join the discussion and vote on comments.

No comments yet. Start the discussion.
Keep exploring

More from this topic

More in OpenClaw