The Agent Skills Directory

[SAFE]: The skill utilizes the official @modelcontextprotocol/inspector package via npx for interacting with MCP servers, which is a recognized and trusted tool within the MCP ecosystem.\n- [SAFE]: All automation for tool fetching, quality analysis, and grading is performed via local bash scripts that employ jq for robust and secure JSON parsing, avoiding unsafe shell execution patterns.\n- [SAFE]: Selection testing is implemented using intent-based prompts for subagents within a controlled simulation environment, providing a safe framework for measuring tool selection accuracy.\n- [SAFE]: No malicious patterns, such as obfuscated code, hardcoded credentials, unauthorized network exfiltration, or persistence mechanisms, were identified in the skill's scripts or documentation.\n- [SAFE]: The skill processes external tool schemas, presenting a standard surface for indirect prompt injection that is mitigated through structural handling.\n
Ingestion points: retrieves JSON-based tool definitions from user-specified MCP server endpoints or commands.\n
Boundary markers: tool schemas are treated as structured data and interpolated into subagent test templates.\n
Capability inventory: the skill environment permits the execution of local automation scripts and the official MCP inspector CLI.\n
Sanitization: employs jq throughout the pipeline to enforce structural validation and safe data extraction from the ingested schemas.

eval-mcp