agent-browser

Browser automation CLI built for agents (vercel-labs/agent-browser, Apache-2.0): a persistent daemon drives Chrome over CDP, and accessibility-tree snapshots give compact @eN refs to act on — no Playwright, no HTML parsing.

npm i -g agent-browser && agent-browser install   # once; downloads Chrome for Testing
agent-browser doctor                              # when anything misbehaves

The core loop

agent-browser open <url>       # 1. open a page
agent-browser snapshot -i      # 2. see it — interactive elements with @eN refs
agent-browser click @e3        # 3. act on a ref (click, fill, type, select…)
agent-browser snapshot -i      # 4. re-snapshot after ANY page change

Refs go stale the moment the page changes — navigation, submit, re-render, dialog. Re-snapshot before the next ref interaction. When refs fail, fall back to semantic locators (find role button click --name "Submit", find label "Email" fill "…"), then raw CSS selectors.