The Agent Skills Directory

[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface in its evaluation harness.
Ingestion points: The script scripts/evaluation.py ingests test questions from user-provided XML files (e.g., evaluation.xml).
Boundary markers: The EVALUATION_PROMPT used in scripts/evaluation.py does not implement boundary markers or instructions to ignore embedded directives around the interpolated question variable.
Capability inventory: The agent loop in the evaluation script is granted access to all tools provided by the MCP server being tested, which may include network access, file system operations, and command execution.
Sanitization: No sanitization or validation of the ingested question content is performed before it is passed to the LLM.
[COMMAND_EXECUTION]: The skill provides tools for executing local commands as part of its primary testing functionality.
Evidence: The scripts/connections.py and scripts/evaluation.py files facilitate the execution of local MCP servers via standard input/output (stdio) transports using subprocess-like logic.
Context: This functionality is the primary intended purpose of the skill, enabling developers to run and evaluate the code they are developing.
[EXTERNAL_DOWNLOADS]: The skill directs the agent to fetch documentation from official external sources.
Evidence: SKILL.md instructs the agent to use WebFetch to load the MCP specification and SDK documentation from the official modelcontextprotocol.io domain and its associated GitHub repositories.
Context: These downloads target well-known and official resources required for the development of MCP servers.

mcp-builder