agent-browser · Openclaw Master Skills

Navigate to a URL, capture an interactive accessibility tree snapshot with JSON output, and click or fill elements using their ref identifiers
Run isolated parallel sessions with separate state persistence to test multi-user workflows such as admin and user contexts

JSON snapshot data containing accessibility tree text and element refs with role and name
JSON responses from get commands (text, html, value, attr, title, url, count)
JSON responses from is commands (visible, enabled, checked)
Saved files: screenshots (.png), PDFs (.pdf), state files (e.g., auth.json)

Not intended for visual inspection or screenshot-based analysis; use the built-in browser tool for that
Does not support browser extension integration; use the built-in browser tool for that
Refs from snapshots become stale after page changes; re-snapshot before interacting following navigation or DOM updates
Confirm every snapshot command includes the -i and --json flags so output is parseable and focused on interactive elements
Verify a re-snapshot is performed after any action that changes the page (open, click, fill, press Enter, reload)
Ensure sessions are named via --session or AGENT_BROWSER_SESSION when isolating concurrent browser contexts to avoid state collision
Confirm a wait command (e.g., wait --load networkidle) is used after navigation or dynamic content updates before snapshotting

AI 搜索