The Agent Skills Directory

[COMMAND_EXECUTION]: The skill contains Python scripts (scripts/evaluation.py and scripts/connections.py) that facilitate the execution of arbitrary local commands. This functionality is intended to allow the developer to launch and test MCP servers via the stdio transport, but it allows for the execution of any system command provided as an argument to the scripts.
[EXTERNAL_DOWNLOADS]: The SKILL.md file contains instructions for the AI agent to fetch documentation and protocol specifications from external URLs, specifically modelcontextprotocol.io and GitHub repositories associated with the Model Context Protocol. These are well-known technology domains and are used to gather necessary technical context for the development task.
[PROMPT_INJECTION]: The evaluation harness in scripts/evaluation.py is vulnerable to indirect prompt injection.
Ingestion points: The script reads and parses task questions from an external XML file (e.g., evaluation.xml or scripts/example_evaluation.xml).
Boundary markers: The agent loop utilizes a system prompt (EVALUATION_PROMPT), but the questions parsed from the XML are interpolated directly into the message chain without specific delimiters or instructions to ignore embedded commands.
Capability inventory: The agent configured in scripts/evaluation.py has the capability to call any tool exposed by the connected MCP server, which could include file system access, network requests, or further command execution depending on the server's implementation.
Sanitization: There is no evidence of sanitization or escaping of the input questions retrieved from the XML before they are processed by the LLM.

mcp-builder