The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The script scripts/evaluation.py reads test questions from user-provided XML files and injects them directly into the LLM conversation without sanitization or strict boundary markers.
Ingestion points: The parse_evaluation_file function in scripts/evaluation.py extracts question text from XML <qa_pair> elements.
Boundary markers: No specific delimiters or instructions are used to wall off the question content, making it possible for a malicious test question to attempt to override the evaluation agent's instructions.
Capability inventory: The evaluation agent has access to all tools exposed by the MCP server being tested, which may include file system access or network operations depending on the server's design.
Sanitization: The raw text from the XML is appended directly to the message history.
[COMMAND_EXECUTION]: The skill facilitates the execution of local commands to initialize MCP servers via the standard input/output (stdio) transport mechanism.
Evidence: scripts/connections.py implements the MCPConnectionStdio class which utilizes the mcp.client.stdio.stdio_client to spawn subprocesses based on a provided command and arguments.
Logic: This is a fundamental component of the Model Context Protocol for local integrations, where the client is responsible for launching and managing the server process.

mcp-builder