The Agent Skills Directory

[SAFE]: The skill serves as a technical reference and development toolkit. All code provided in the scripts and references is dedicated to implementing and testing MCP servers using official libraries.\n- [COMMAND_EXECUTION]: The evaluation script (scripts/evaluation.py) supports a stdio transport mode that executes a user-provided command to launch a local MCP server. This allows for automated testing of server functionality but relies on the user providing a safe command string for the local environment.\n- [EXTERNAL_DOWNLOADS]: The implementation guide (SKILL.md) instructs the agent to fetch the official MCP specification and SDK documentation from modelcontextprotocol.io and the official GitHub repositories for the protocol. These are recognized as trusted and well-known sources for development documentation.\n- [DATA_EXFILTRATION]: The evaluation harness in scripts/evaluation.py ingests data from XML evaluation files and incorporates tool outputs from external MCP servers into the agent loop. This is a potential surface for indirect prompt injection, which is mitigated by delimiter-based boundaries and structured output extraction.\n
Ingestion points: XML evaluation files (parse_evaluation_file) and MCP tool responses (connection.call_tool).\n
Boundary markers: The agent is instructed to use specific XML-style tags (, , ) to separate internal reasoning from final outputs.\n
Capability inventory: The script possesses the capability to execute local commands via stdio_client and perform network requests to the Anthropic API and remote MCP servers.\n
Sanitization: The script isolates the final response using regular expression extraction of content within the designated tags, providing a layer of separation from potentially malicious tool output.

mcp-builder