The Agent Skills Directory

[COMMAND_EXECUTION]: The evaluation harness in scripts/evaluation.py and scripts/connections.py supports the stdio transport, which executes local MCP servers as subprocesses. This allows the tool to run and test integrations directly on a developer's machine using user-provided commands and arguments.
[INDIRECT_PROMPT_INJECTION]: The automated agent loop in evaluation.py processes external data from evaluation XML files and results from third-party MCP tools, creating a surface for indirect prompt injection.
Ingestion points: Evaluation question strings and MCP tool execution outputs.
Boundary markers: The system prompt uses <summary>, <feedback>, and <response> XML tags to structure agent output.
Capability inventory: The script can spawn local processes and perform network requests to the Anthropic API.
Sanitization: Tool results are converted to strings or JSON without specific adversarial filtering before being sent to the model.
[EXTERNAL_DOWNLOADS]: The implementation guides instruct the agent to use the WebFetch tool to retrieve documentation and README files from official Model Context Protocol repositories on GitHub. These are references to well-known, trusted sources for the protocol.

mcp-builder