The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/evaluation.py script and scripts/connections.py module provide the capability to launch local MCP servers as subprocesses using the stdio transport. This behavior is triggered by user-supplied command-line arguments and is essential for testing locally developed servers.
[EXTERNAL_DOWNLOADS]: The SKILL.md file and implementation guides recommend fetching the latest protocol specifications and SDK README files from modelcontextprotocol.io and its official GitHub repository. These sources are official and well-known within the technical domain of the skill.
[PROMPT_INJECTION]: The evaluation harness in scripts/evaluation.py is susceptible to indirect prompt injection because it processes tool outputs and external XML-based questions. A malicious tool output or evaluation task could theoretically contain instructions intended to override the evaluator agent's behavior. (1) Ingestion points: Untrusted data enters the agent context through the evaluation XML file and tool responses returned by the connection.call_tool method in scripts/evaluation.py. (2) Boundary markers: There are no explicit boundary markers or 'ignore embedded instructions' warnings used when tool outputs are appended to the agent's message history. (3) Capability inventory: The script can execute any tool available on the connected MCP server, which typically includes data retrieval and, depending on the server, potentially destructive operations. (4) Sanitization: The script does not sanitize or escape tool outputs before they are processed by the LLM.

mcp-builder