browser-use

Take an accessibility snapshot of the page to get element references, then call click or fill actions with those observed references. For multi-field forms, use the fill_form tool with a list of field values; filling never submits the form automatically.

Start your existing development server, then navigate the browser to its loopback HTTP URL. Direct file: navigation is unsupported, so standalone HTML needs a static server bound to loopback serving only the project directory.

In Workbench it uses the browser selected in Browser settings, including an existing Chrome connection with your chosen profile, so cookies and site logins are shared. Standalone Pi uses an isolated headless Chrome instance without access to another browser's profile.

References go stale after navigation, element removal, or eviction, so take a fresh snapshot after a stale-reference error. A fresh reference still fails if the element is hidden, disabled, or covered; inspect the blocking UI instead of repeating the click.

Browser automation suits tasks requiring the page UI, but a connector or API is preferred for semantic tasks that do not need page interaction. CAPTCHA and consent challenges are reported rather than bypassed, and cross-origin frames remain unavailable in snapshots.