The Agent Skills Directory

COMMAND_EXECUTION (MEDIUM): The scripts/connections.py file includes an MCPConnectionStdio class that uses the stdio_client to launch local processes. This utility allows an agent to execute arbitrary commands and arguments on the host system. This is a high-risk capability if the command parameters are influenced by untrusted input.
PROMPT_INJECTION (LOW): The skill templates in the examples/ directory are susceptible to Indirect Prompt Injection (Category 8) due to the ingestion of external API data without output sanitization or boundary markers.
Ingestion points: External data enters the context via fetch_from_api and fetch_user in the examples/ directory scripts.
Boundary markers: No specific delimiters or 'ignore instructions' warnings are included in the response templates within the example code.
Capability inventory: The skill provides subprocess execution via scripts/connections.py and network access capabilities.
Sanitization: Input validation is performed via Pydantic and Zod schemas, but there is no evidence of sanitizing external data before it is returned to the agent context.
EXTERNAL_DOWNLOADS (LOW): SKILL.md contains instructions for the agent to download protocol specifications and SDK details from modelcontextprotocol.io and github.com. These are trusted sources per standard guidelines, but the behavior constitutes an external dependency for the skill logic.

mcp-builder