The Agent Skills Directory

[COMMAND_EXECUTION]: The testing harness included in the skill executes local servers as part of its intended functionality.
Evidence: The scripts/evaluation.py and scripts/connections.py files facilitate the execution of local MCP servers via standard input/output (stdio).
Detail: This allows developers to pass commands (e.g., python, node) and script arguments through CLI parameters to test their server implementations locally.
[PROMPT_INJECTION]: The evaluation harness processes responses from external MCP tools, which constitutes an indirect prompt injection surface.
Ingestion points: scripts/evaluation.py (lines 98-100) reads tool results from the connected server.
Boundary markers: The harness utilizes the structured tool_result message type from the Anthropic SDK to delimit tool outputs within the conversation context.
Capability inventory: The agent can invoke any tool exposed by the connected MCP server, including those involving file system access or network operations defined by the server being tested.
Sanitization: Tool outputs are serialized to strings or JSON (lines 99-100) before being interpolated into the messages sent back to the LLM.
[EXTERNAL_DOWNLOADS]: The documentation provides links to official protocol resources and SDKs.
Evidence: SKILL.md and reference guides include URLs to modelcontextprotocol.io and the modelcontextprotocol organization on GitHub.
Detail: These references are to well-known, trusted organizations and are used strictly for retrieving legitimate documentation and configuration samples.

mcp-builder