The Agent Skills Directory

[COMMAND_EXECUTION]: The evaluation scripts (scripts/evaluation.py and scripts/connections.py) allow for the execution of local system commands using the MCP stdio transport. This is the intended functionality for testing locally developed MCP servers.
[EXTERNAL_DOWNLOADS]: The documentation suggests fetching SDK and protocol information from the official Model Context Protocol GitHub repositories and website. These are reputable and trusted sources for technical documentation.
[PROMPT_INJECTION]: The evaluation harness implements an agent loop that processes tool outputs, creating a surface for potential indirect prompt injection.
Ingestion points: Tool results are retrieved in scripts/evaluation.py via the connection.call_tool method.
Boundary markers: The EVALUATION_PROMPT in scripts/evaluation.py uses XML-like tags (<summary>, <feedback>, <response>) to structure the model's output.
Capability inventory: The agent is permitted to invoke any tools provided by the MCP server under test, which may involve file system or network operations.
Sanitization: Tool outputs are integrated into the message history as strings without explicit sanitization, relying on the model's instructions and internal safety layers.
[SAFE]: Sensitive information like API keys is managed through environment variables rather than being hardcoded in the skill's scripts or documentation.

mcp-builder