The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/evaluation.py and scripts/connections.py files implement the stdio transport for MCP servers. This involves spawning a local subprocess using command-line arguments (e.g., -c python -a server.py) provided by the user. This is the standard mechanism for local MCP integration and is controlled by the user.
[INDIRECT_PROMPT_INJECTION]: The scripts/evaluation.py script reads test questions from an external XML file and includes them in the prompt context for an LLM during the evaluation phase. This provides a surface where untrusted data could influence agent behavior, though this is the intended function of a testing harness.
Ingestion points: Questions are read from an XML file (e.g., evaluation.xml) in scripts/evaluation.py via parse_evaluation_file.
Boundary markers: The questions are wrapped in a user message within the agent_loop function.
Capability inventory: The evaluator agent has access to the tools exposed by the MCP server being tested.
Sanitization: The script performs no specific sanitization of the question text before passing it to the LLM.

mcp-builder