unstucklisted
Install: claude install-skill velmlabs/unstuck
# Unstuck: hand a stuck browser step to a human
Unstuck pauses the agent, sends a person a link to the live browser, lets them tap and type on it from their phone, and returns control when they tap **Done**. The page, cookies and session stay exactly where they were.
## When to call it
Call a takeover only when a human is the only way forward:
- a CAPTCHA, "I'm not a robot" check or bot challenge;
- a login that needs a password you don't have, or a passkey;
- a 2FA, SMS or email code;
- a page you can't make sense of after one honest retry (an unexpected consent screen, a broken form).
Don't call it for things you can do yourself (scrolling, closing a normal pop-up, retrying a slow page), and never to get past a site's rules: the human is asked to do the step, not to help you abuse the site.
When you need a **decision** rather than a page (which of two flights, whether to send a message, a detail only the user knows), don't take over a page: ask a question (`ask_question` over MCP, `ask()` in code). See *Ask the person a question*.
## Choose the integration
1. **MCP tools available** (`request_human_takeover`, `start_human_takeover` and `ask_question`, from the `unstuck` MCP server): use them. Go to *MCP*.
2. **You write the automation code** (TypeScript or Python with Playwright, Puppeteer, browser-use, Stagehand): use the SDK. Go to *Code*.
3. **You drive the browser one shell command at a time** (agent-browser, or any CLI that talks to Chrome over CDP): use the `u