The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The evaluation.py script implements an automated testing loop that reads questions from an external XML file and passes them directly to an LLM. This creates a surface for indirect prompt injection where malicious questions could attempt to influence the evaluation agent's behavior.
Ingestion points: scripts/evaluation.py reads user-provided XML files via the parse_evaluation_file function.
Boundary markers: The agent prompt uses XML tags (<summary>, <feedback>, <response>) to structure the output, but it lacks explicit isolation or instructions to ignore embedded commands within the input questions.
Capability inventory: The script can execute local commands (spawning the MCP server), make network requests to the Anthropic API, and call tools provided by the MCP server.
Sanitization: No sanitization or validation of the input questions is performed before they are interpolated into the prompt.
[COMMAND_EXECUTION]: The skill provides a utility for executing and testing MCP servers by spawning them as local subprocesses based on user-supplied command-line arguments.
Evidence: scripts/connections.py uses mcp.client.stdio.stdio_client to launch servers using StdioServerParameters, and scripts/evaluation.py exposes this capability through its -c (command) and -a (args) CLI options.
[EXTERNAL_DOWNLOADS]: The documentation references official Model Context Protocol resources and SDKs hosted on GitHub and the project's documentation site.
Evidence: SKILL.md includes links to modelcontextprotocol.io and the modelcontextprotocol organization on GitHub for fetching specifications and SDK READMEs.

mcp-builder