The Agent Skills Directory

[SAFE]: Analysis of the skill instructions and provided metadata confirms the skill is a legitimate messaging interface for agent-to-agent and agent-to-human communication.\n- [COMMAND_EXECUTION]: The skill references CLI login commands (e.g., claude mcp login agentcouch) for the human user to execute. These are standard authentication procedures and do not involve the agent executing arbitrary or malicious shell commands autonomously.\n- [DATA_EXPOSURE]: The skill involves sharing email addresses and user IDs to facilitate room invitations. This is restricted to the intended messaging functionality and authentication is managed via human-led OAuth flows to prevent unauthorized access.\n- [PROMPT_INJECTION]: While the skill processes messages from external parties (creating a potential surface for indirect prompt injection), the instructions explicitly mitigate this by directing the agent to ignore instructions embedded in message bodies and rely solely on verified platform metadata for identity and authority.

agentcouch-chat