The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads a Docker Compose configuration file from a third-party GitHub repository and points users to binary releases.
Evidence: wget https://raw.githubusercontent.com/xpzouying/xiaohongshu-mcp/main/docker/docker-compose.yml in SKILL.md.
Evidence: Instructions to fetch binary URLs from https://api.github.com/repos/xpzouying/xiaohongshu-mcp/releases/latest.
[REMOTE_CODE_EXECUTION]: The skill guides the user to execute external code in the form of Docker containers and downloaded binaries from an unverified source.
Evidence: docker compose up -d pulls and runs images from xpzouying/xiaohongshu-mcp.
Evidence: Step 2, Way 2: 'From GitHub Releases download... binary mode needs Chrome installed locally.'
[COMMAND_EXECUTION]: The skill uses shell commands for network checks, service management, and tool configuration.
Evidence: curl -so /dev/null http://localhost:18060/mcp && echo "running" || echo "not running".
Evidence: claude mcp add xiaohongshu --transport http <地址>.
[INDIRECT_PROMPT_INJECTION]: The skill reads and modifies sensitive configuration files, which could be an attack surface if external data sources are processed.
Ingestion points: Reading and checking ~/.claude/settings.json and .claude/settings.json in SKILL.md.
Boundary markers: Absent.
Capability inventory: curl, wget, docker, claude (in SKILL.md).
Sanitization: Absent; the skill directly interpolates addresses into configuration files.

setup-xhs-mcp