The Agent Skills Directory

COMMAND_EXECUTION (SAFE): The skill uses subprocess.run within scripts/health_check.py to verify binary presence using the which command. This is implemented using a list of arguments, which prevents shell injection attacks.
DATA_EXPOSURE (SAFE): The scripts access the local ~/.claude.json file to identify configured MCP servers. While this file can contain API tokens, the skill processes this information only for local management and does not contain any network communication or exfiltration logic.
INDIRECT_PROMPT_INJECTION (SAFE): This skill has an ingestion surface for data from local configuration files.
Ingestion points: Reads ~/.claude.json in scripts/health_check.py and scripts/mcp_manager.py.
Boundary markers: None; the script assumes a valid JSON structure from the local config.
Capability inventory: Limited to file read/write and existence checks via subprocess.run(['which', ...]).
Sanitization: No specific sanitization of the command strings retrieved from the config, but the impact is minimized by using non-shell execution.

mcp-manager