Browser
You can drive a real Chromium running inside this container. With this skill
active, the browser core starts automatically and you have browser_* MCP
tools attached over CDP to the live shared session. If the user opens the
Browser pane, they see the same browser and can log in by hand; you inherit
that session's cookies without handling credentials.
This skill is for the Live Agent Browser only. For one-off public
screenshots, recordings, or cookie-authenticated headless recipes, use
/workspace/scripts/browser.mjs instead.
How to work
Use a hybrid perception loop:
browser_navigate, thenbrowser_snapshot. The snapshot is the default way to read structure, text, roles, and element refs.- Take a screenshot when the task is visual: layout verification, charts, maps, canvas, image content, custom widgets, anything opaque or missing from the snapshot, and after actions with visible consequences.
- Act through refs when the snapshot gives reliable targets. Use screenshot-grounded coordinate interaction when the DOM is misleading.
- Observe again after each action with
browser_snapshot,browser_take_screenshot, orbrowser_wait_for. Do not assume an action worked.
Keep screenshot cost bounded: viewport screenshots by default, JPEG/moderate quality when options are available, full-page screenshots only when needed. The viewport is 1366x768 and maps 1:1 to OS-level fallback coordinates.
Pacing
Wait for navigation/load states before acting. Use browser_wait_for when
content is dynamic. Take one meaningful action, then observe. Avoid blind
multi-action bursts, especially on authenticated sites.
Logging in
This browser exits via the container's datacenter IP, so strict providers may show verification challenges. If a task needs a login or challenge response:
- Ask the user to open the Browser pane and log in by hand.
- Wait until they say the login is complete.
- Continue with the
browser_*tools against the shared session.
Never type the user's credentials yourself.
OS-level input fallback
Use the browser_* tools first. If a site visibly swallows or rejects CDP
input, fall back to X-server input with coordinates from a screenshot:
sh /workspace/.browser-gui/human-input.sh move 640 420
sh /workspace/.browser-gui/human-input.sh click 640 420
sh /workspace/.browser-gui/human-input.sh type "text to type"
sh /workspace/.browser-gui/human-input.sh key Tab Return
sh /workspace/.browser-gui/human-input.sh scroll -3
Coordinates map to the browser window at 0,0 on the 1366x768 display.
Write policy
Reading (search, timelines, messages) is fine on your own. Before any public or irreversible write - posting, replying, DMing, following, buying, or changing settings - say exactly what you're about to do and get the user's confirmation first. They can watch and stop you in the Browser pane.