← ClaudeAtlas

agent-hands-corelisted

Core agent-hands usage guide. Read before running any agent-hands command.
yassiEmp/agent-hands · ★ 4 · AI & Automation · score 68
Install: claude install-skill yassiEmp/agent-hands
# agent-hands core Human-rate mouse and keyboard for a real browser. ## The loop, and why you stay inside it ```bash agent-hands browsers # what exists, what is reachable # more than one reachable? ASK THE USER which to drive. agent-hands snapshot --browser edge # see the page: @e1 @e2 @e3 agent-hands click --ref @e5 --browser edge ``` `snapshot` mints its own refs now, so this works on the user's own browser and not only on a pooled session. You never need a raw CDP script to see a page. **Do not write one.** A hand-rolled CDP or Playwright script against the same browser costs three things at once: - **It dispatches input instantly.** That is the signal behaviour-scoring sites grade. This tool exists to avoid it; bypassing it forfeits the whole point. - **It opens its own connection.** The relay holds the single approved socket, so a second connection re-prompts the user, and on a browser holding their live session that prompt is the only thing standing between an agent and their accounts. - **It leaves no trace.** `agent-hands audit` records every authorisation the relay grants and every client that used it, so a raw socket on a `--browser` target shows up there as an absence. That covers the relayed lane only: a `--cdp` port and a pooled `--session` need no approval, so nothing authorises and nothing is logged. `agent-hands audit` tells you which lane you are on. Measured against `bot-detector.rebrowser.net`: a snapshot, a scr