The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes functionality to launch and interact with local MCP servers.
Evidence: scripts/connections.py utilizes the stdio_client from the mcp library, which spawns subprocesses based on user-provided commands and arguments (e.g., python server.py). This is a core requirement for testing locally developed MCP servers.
[EXTERNAL_DOWNLOADS]: The instructions direct the agent to fetch documentation and SDK information from external sources to ensure the latest protocol standards are followed.
Evidence: SKILL.md fetches content from modelcontextprotocol.io and the official modelcontextprotocol organization on GitHub. These are well-known, trusted sources for this specific technology domain.
[INDIRECT_PROMPT_INJECTION]: The evaluation script processes outputs from external tools, which represents a potential attack surface if those tools return malicious instructions.
Ingestion points: scripts/evaluation.py (via tool_result = await connection.call_tool(...)) and SKILL.md (via documentation fetching).
Boundary markers: The EVALUATION_PROMPT in scripts/evaluation.py uses XML-style tags (<summary>, <feedback>, <response>) to structure the agent's response, providing some separation from tool outputs.
Capability inventory: The skill can execute local commands (via stdio transport), perform network requests (via SSE/HTTP transports), and interact with the Anthropic API.
Sanitization: The evaluation script directly serializes tool outputs into the message history without explicit instruction-filtering, though it relies on the model's inherent ability to distinguish between tool results and user instructions.

mcp-builder