The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/connections.py and scripts/evaluation.py scripts implement a stdio transport mechanism that uses subprocesses to launch and interact with local MCP server binaries or scripts provided by the user via CLI arguments.
[REMOTE_CODE_EXECUTION]: The scripts/evaluation.py harness executes an agentic loop using the Anthropic API to dynamically call tools on local or remote MCP servers. The tools called are determined at runtime by the AI model based on the provided test questions.
[EXTERNAL_DOWNLOADS]: The SKILL.md file directs the agent to fetch documentation, sitemaps, and SDK README files from the official modelcontextprotocol GitHub repositories and website to aid in the development process.
[PROMPT_INJECTION]: The scripts/evaluation.py script possesses an indirect prompt injection surface by ingesting untrusted tool outputs directly into the agent's context. Evidence: 1. Ingestion points: MCP tool results within the agent loop; 2. Boundary markers: The system prompt uses XML tags (<summary>, <feedback>, <response>) to structure agent output; 3. Capability inventory: The harness allows for tool execution and subprocess command execution; 4. Sanitization: No specific filtering or sanitization of tool outputs is performed before they are passed to the model.

mcp-builder