The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/evaluation.py uses the stdio_client from the mcp library to manage subprocesses. This functionality is intended to allow the agent to test local MCP servers by executing commands provided through CLI arguments.
[EXTERNAL_DOWNLOADS]: The workflow in SKILL.md directs the agent to fetch the MCP protocol specification and SDK documentation from official domains, specifically modelcontextprotocol.io and the modelcontextprotocol organization on GitHub.
[PROMPT_INJECTION]: The evaluation framework provided in scripts/evaluation.py creates a surface for indirect prompt injection by processing external data from XML files and server outputs.
Ingestion points: Evaluation questions are parsed from user-provided XML files in scripts/evaluation.py, and tool execution results are retrieved from the MCP server being evaluated.
Boundary markers: The harness uses XML tags such as <summary>, <feedback>, and <response> to structure the agent's output, but it does not implement strong delimiters to separate untrusted input data from the system prompt.
Capability inventory: The environment supports subprocess execution (via stdio_client) and network communication via SSE or HTTP transport methods.
Sanitization: Data parsed from the XML files and tool results are integrated into the LLM context without validation or escaping mechanisms.
[DATA_EXFILTRATION]: The scripts/evaluation.py script utilizes the xml.etree.ElementTree library to parse evaluation XML files. This library is vulnerable to XML External Entity (XXE) attacks, which could potentially be used to expose sensitive local files if a malicious XML file is supplied as input.

mcp-server