Agent Browser Skill

Two browser automation options are available, sharing the same Chrome for Testing v148 binary:

1. OpenClaw browser tool — built-in plugin, managed by gateway, auto-started
2. agent-browser CLI (v0.26.0) — standalone, extra features (batch, auth vault, visual diff, chat)

Rule of thumb: Public websites → OpenClaw browser tool. Logged-in sites → Tandem Browser profile="user". Complex automation → agent-browser CLI via exec.

Quick Reference

Status

browser doctor           # Check health
browser status           # Running state
browser start            # Launch (auto-start if not running)
browser stop             # Shutdown

Navigate & Read

browser navigate url="https://example.com"
browser snapshot              # Accessibility tree with @ref IDs
browser snapshot refs="aria"  # Playwright aria refs (more stable)
browser screenshot            # Visual capture

Act (use snapshot refs)

browser act request={kind:"click", ref:"@e1"}
browser act request={kind:"fill", ref:"@e3", text:"value"}
browser act request={kind:"type", ref:"@e3", text:"value slowly"}
browser act request={kind:"press", key:"Enter"}
browser act request={kind:"select", ref:"@e2", values:["option1"]}
browser act request={kind:"hover", ref:"@e5"}
browser act request={kind:"wait", timeMs:2000}

Tab Management

browser open url="...", label="my-tab"   # Open with named label
browser focus targetId="my-tab"           # Switch to tab
browser close targetId="t1"               # Close tab
browser tabs                              # List all open tabs

Agent-Browser CLI (via exec)

agent-browser open https://...
agent-browser snapshot -i -c
agent-browser click @e2
agent-browser fill @e3 "text"
agent-browser screenshot --annotate
agent-browser keyboard type "text"
agent-browser auth save <name> --url <url> --username <u> --password <p>
agent-browser diff snapshot
agent-browser batch ["open url" "snapshot -i" "click @e1"]

Operating Loop

1. Check status — browser status or browser doctor if health unknown
2. Navigate — browser navigate url="<target>"
3. Read — browser snapshot to get interactive elements with @ref IDs
4. Act — Use refs from snapshot for click/fill/press/select
5. Re-snapshot — After any navigation or UI change, snapshot again (refs become stale)
6. Recover — If ref fails, snapshot once more; if still broken, report as manual action
7. Profile pick — Omit profile for public sites; use profile="user" for logged-in Tandem sessions

Tips

• Use refs="aria" for more stable element references that survive page re-renders
• Label task tabs at open time: label="research"
• agent-browser chat "..." for AI-driven natural language browser tasks
• agent-browser --headed to show browser window for debugging
• agent-browser --profile Default to reuse Chrome login state