Browser automation

Browser tools use Playwright to inspect pages, capture evidence, and run reproducible browser workflows. The public surface is deliberately small.

Tools

Tool	Purpose
browser_session	Start, list, close, or clean up persistent browser sessions; optionally reuse a profile or storage state.
browser_snapshot	Read bounded page text, page/network errors, and interactive elements with short refs such as e1; optionally capture a screenshot.
browser_act	Run structured navigation, click, fill, select, key, wait, and multi-page actions using snapshot refs or CSS selectors.
browser_run_script	Run a complete Python Playwright script when the high-level action set is insufficient.

All browser tools accept optional machine. Browser dependencies must already be installed on the selected controller or worker; installation is performed with ordinary shell commands such as python -m playwright install chromium.

Common flows

For interactive work, call browser_session(action="start", url=...), then browser_snapshot. The snapshot returns short references such as e1 and e2; pass those refs directly to browser_act, for example {"action": "click", "target": "e1"} or {"action": "fill", "target": "e2", "value": "..."}. Re-snapshot after navigation because element refs are page-state references, not permanent selectors.

For ordinary inspection and screenshots, prefer browser_session plus browser_snapshot; the snapshot can return bounded visible text and save a screenshot. Use browser_run_script for JavaScript evaluation, custom capture/PDF logic, or interactions not represented by browser_act.

Keep scripts bounded, set explicit timeouts, save artifacts under the workspace, and avoid entering credentials unless the environment is dedicated to the task.