The Agent Skills Directory

[COMMAND_EXECUTION]: The evaluation harness and connection library support launching local processes as MCP servers.
Evidence in scripts/evaluation.py and scripts/connections.py: The stdio transport mode uses the mcp SDK to execute commands (e.g., python server.py) provided by the user via command-line arguments to facilitate local testing and development.
[EXTERNAL_DOWNLOADS]: The skill documentation directs the agent to retrieve the latest protocol specifications and SDK documentation from official sources.
Evidence in SKILL.md: Instructs the agent to use WebFetch to load documentation from https://modelcontextprotocol.io and raw README files from the modelcontextprotocol organization on GitHub. These sources are established for the protocol and are used exclusively for retrieving informational markdown files.
[INDIRECT_PROMPT_INJECTION]: The skill provides an evaluation script that processes external data files, which creates a surface for indirect prompt injection during testing.
Ingestion points: scripts/evaluation.py parses XML files (e.g., scripts/example_evaluation.xml) containing test questions that are passed directly into the LLM's prompt.
Boundary markers: The system prompt in scripts/evaluation.py defines specific output tags (<summary>, <feedback>, <response>) to structure the model's output and mitigate instruction leakage, though the input questions are not encapsulated in unique delimiters.
Capability inventory: The agent being evaluated has access to the tools exposed by the MCP server being tested via the connection.call_tool method.
Sanitization: The questions extracted from the XML are not sanitized or escaped before interpolation into the prompt, as the script is intended for use with developer-controlled test suites.

mcp-builder