chrome-automation

A browser-automation skill that drives tasks in the user's real Chrome session through the agent-browser CLI. It can navigate pages, fill forms, click buttons, take screenshots, extract data, and replay recorded workflows, all inside the user's existing browser rather than a fresh Playwright instance. Its core principle is to reuse the single Chrome process the user already has open, which preserves cookies, login sessions, and page state, an important advantage for marketing platforms like social dashboards, ad managers, and CMS tools that require login. As a prerequisite, agent-browser must be installed and Chrome must have remote debugging enabled.

The recommended pattern always starts by listing existing tabs with agent-browser and the tab list command; if the target page is already open it switches to that tab, otherwise it opens the URL. The auto-connect flag discovers Chrome with remote debugging; for Chrome 144 and later WebSocket-only endpoints where discovery URLs return 404, the guidance shows connecting via npx agent-browser at latest to a devtools browser WebSocket endpoint on 127.0.0.1 port 9222. Interaction relies on the snapshot command to enumerate interactive elements as refs like at-e1 and at-e2, which are then used with click, fill, keyboard inserttext, press, scroll, wait, eval, and screenshot. A step-type mapping table distinguishes standard inputs, which use fill, from contenteditable rich-text editors, which need focus first and then keyboard inserttext, and it stresses that refs are invalidated when the page changes, so re-snapshotting after navigation, form submission, or dynamic loads is required.

It can replay recordings exported from the Chrome DevTools Recorder in JSON (recommended, readable progressively with jq), @puppeteer/replay JS, or Puppeteer JS formats, matching the recording's aria selectors against snapshots and falling back to text selectors, CSS hints, then screenshots, while avoiding brittle ember IDs, numeric IDs, and exact XPaths. A dedicated section handles iframe-heavy sites such as LinkedIn and Gmail, where the snapshot command cannot penetrate iframes: workarounds include using eval on same-origin iframe content, keyboard input for focused elements, and reading full page text. Setup details live in references/agent-browser-setup.md.