The Agent Skills Directory

[COMMAND_EXECUTION]: The 'mcp_connect' tool enables arbitrary system command execution via the 'stdio' transport. The 'command' and 'args' parameters are passed directly to 'asyncio.create_subprocess_exec' in 'client.py', allowing the agent to launch any binary or script available on the host system without restriction.\n- [EXTERNAL_DOWNLOADS]: The skill facilitates outbound network connections to arbitrary URLs using 'SSETransport' and 'StreamableHTTPTransport'. These transports use the 'httpx' library to connect to unverified external servers, posing a risk of interacting with malicious infrastructure.\n- [DATA_EXFILTRATION]: The connection management logic in 'registry.py' resolves environment variables into HTTP headers or subprocess environments. If directed to a malicious endpoint, sensitive information like API keys or credentials stored in environment variables could be transmitted externally.\n- [REMOTE_CODE_EXECUTION]: The combination of 'stdio' subprocess execution and the ability to define environment variables ('env' parameter) allows for the setup and execution of complex, potentially malicious payloads on the host environment.

mcp-connector