The Agent Skills Directory

Indirect Prompt Injection Surface: The evaluation harness (scripts/evaluation.py) processes external XML files to run automated tests. The content within the <question> tags is interpolated directly into the agent's prompt as a user message. Without strict sanitization or boundary markers, a crafted evaluation file could potentially influence the model's behavior.
Ingestion points: The parse_evaluation_file function in scripts/evaluation.py reads data from user-specified XML files.
Boundary markers: The script does not use specific delimiters or instructions to isolate the question content from the surrounding system prompt.
Capability inventory: The script executes tools on the connected MCP server and communicates with the Anthropic API.
Sanitization: The content from the XML tags is processed as raw text without validation or escaping.
Local Command Execution: The skill includes a test harness designed to launch and interact with local MCP servers via the stdio transport. The scripts/evaluation.py script accepts commands and arguments via CLI flags (such as -c and -a) to execute the server under test. While this is a standard requirement for a test runner, users should ensure that only trusted commands and scripts are provided.
External Resource Access: The implementation guides fetch supplementary documentation from external repositories, such as the official Model Context Protocol GitHub organization. These references are used to provide up-to-date SDK information and best practices during the development workflow.

mcp-builder