The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references official documentation and SDK information from the Model Context Protocol (MCP) website and GitHub repositories for the purpose of guiding development.\n- [COMMAND_EXECUTION]: The evaluation harness in scripts/evaluation.py and scripts/connections.py is capable of launching local MCP server processes using the stdio transport. This is a standard feature for testing Model Context Protocol integrations during the development lifecycle.\n- [INDIRECT_PROMPT_INJECTION]: The evaluation scripts ingest data from external XML files and process results from tool executions, which represents a potential surface for indirect prompt injection.\n
Ingestion points: scripts/evaluation.py (parsing question/answer pairs from XML) and the agent_loop function (receiving outputs from MCP tools).\n
Boundary markers: The harness uses a system prompt (EVALUATION_PROMPT) that defines specific XML tags (<summary>, <feedback>, <response>) to help the model distinguish between its reasoning, feedback, and final answers.\n
Capability inventory: The system can call any tool exposed by the MCP server being tested and execute commands via the stdio connection transport.\n
Sanitization: Input from evaluation files and tool outputs are incorporated into the prompt context without specialized sanitization or escaping mechanisms.

mcp-builder