About this skill

Problem

When scripts must drive a web page, brittle selectors, async loading, and headed-browser assumptions make automation fragile. agent-browser turns a headless browser into a command-driven workflow: open the page, snapshot it, then interact through stable refs. This fits form submission, saved login state, screenshots/PDF export, network observation, and other flows that need CLI output a parser can consume.

How It Works

The core loop is:
- open the target URL;
- run snapshot -i to return elements with refs such as @e1 and @e2;
- interact by ref with clicks, input, waits, and mouse control;
- re-snapshot after navigation or significant DOM changes.
Beyond basic interaction, the skill exposes cookies/storage, network, tabs, frames, dialogs, JavaScript, state management, and parallel --session instances. Add --json for machine-readable output, use --headed for visible debugging, and connect through --cdp when using Chrome DevTools Protocol.

Boundaries

Refs are stable only for the current page load and change after navigation;
prefer fill over type so existing input text is cleared;
add wait after navigation when the page has not finished loading;
recording starts a fresh context while preserving cookies/storage from the current session.

Screenshots

Use Cases

Open a form page in CI, use snapshot refs to submit data, and emit JSON output.
Reuse saved cookies and storage while running parallel browser sessions for batch tasks.
Re-snapshot after navigation, then locate inputs, buttons, and tables by stable refs.
Debug headless failures with --headed, then stabilize page state using --json and wait.

Best For

Backend engineers who need web actions as parseable CLI commands
Automation engineers stabilizing form submissions and page state in CI
Crawler or data engineers reusing login state and parallel sessions
Frontend toolchain developers debugging headless interactions and screenshots/PDF