The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements an evaluation harness that processes external, potentially untrusted data to test MCP servers.
Ingestion points: The scripts/evaluation.py script reads question and answer pairs from a user-supplied XML file and processes tool outputs from the connected MCP server.
Boundary markers: The system prompt in scripts/evaluation.py defines XML tags (<summary>, <feedback>, <response>) for agent output structure, but does not provide delimiters or "ignore embedded instructions" warnings for the input questions.
Capability inventory: The script has the ability to execute any tool exposed by an MCP server and can spawn arbitrary local subprocesses when using the stdio transport.
Sanitization: The script does not perform sanitization, escaping, or validation of the content within the evaluation questions before passing them to the LLM.
[EXTERNAL_DOWNLOADS]: The skill instructions guide the agent to fetch documentation and specification files from official and well-known sources.
Fetches the Model Context Protocol specification from the official domain modelcontextprotocol.io.
Downloads SDK documentation and README files from the official modelcontextprotocol organization on GitHub.
[COMMAND_EXECUTION]: The evaluation scripts facilitate the execution of local MCP servers as subprocesses for testing purposes.
scripts/evaluation.py and scripts/connections.py allow for launching local processes via the stdio transport using commands and arguments provided at runtime.
[DYNAMIC_EXECUTION]: The evaluation harness uses an agent loop that dynamically determines and executes tool calls.
scripts/evaluation.py interprets LLM responses to call tools on the connected MCP server dynamically during the execution of evaluation tasks.

mcp-builder