Unstuck Docs GitHub

Agent skill

The Unstuck skill teaches your AI agent when to hand a browser step to a person and how to do it: which calls to make, what to put in the reason, and what to do with each result. Install it once and your agent uses Unstuck whenever it meets a CAPTCHA, a login, a 2FA code or a page it can't get past.

Claude Code

Add the Unstuck plugin. It brings the skill and the Unstuck MCP server (the request_human_takeover tool) in one step:

/plugin marketplace add velmlabs/unstuck
/plugin install unstuck@unstuck

Start Chrome with a debugging port so the agent has a browser to hand over:

google-chrome --remote-debugging-port=9222 --user-data-dir="$HOME/.unstuck-chrome"

On a Mac, or with Claude Desktop or Cursor, see the MCP guide.

Then just ask for the task, for example: "Open the supplier portal at portal.example.com and download this month's invoice." When the agent reaches the login or a CAPTCHA, it calls Unstuck, you get a link, and it carries on once you tap Done.

Any agent that reads skills

Codex, Cursor, Claude Desktop and other agents that support skill files can use the same file. Download it into the agent's skills folder:

mkdir -p ~/.claude/skills/unstuck
curl -fsSL https://unstuck.sh/skills/unstuck/SKILL.md -o ~/.claude/skills/unstuck/SKILL.md

Use your agent's own skills folder if it isn't ~/.claude/skills. The skill tells the agent to use the MCP tool when it has one and the TypeScript or Python SDK when it writes the automation code, so also add the MCP server or install the SDK.

Without an API key, takeover links open on the agent's own machine only. To open them on your phone from anywhere, and to get phone alerts, create a key on the API keys page and set it where the agent runs:

export UNSTUCK_API_KEY=uk_live_…

What the skill tells your agent

You can read the whole skill at unstuck.sh/skills/unstuck/SKILL.md.