The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches README files and protocol specifications from official, well-known sources, including the modelcontextprotocol GitHub organization and its official documentation site. These resources are used to provide the agent with the latest protocol standards and SDK usage patterns.
[COMMAND_EXECUTION]: The scripts/evaluation.py and scripts/connections.py files provide functionality to execute local commands for starting MCP servers when using the stdio transport. This execution is an intended core feature triggered by specific user-provided arguments for the purpose of automated testing.
[PROMPT_INJECTION]: The evaluation loop in scripts/evaluation.py processes untrusted data from user-provided XML files and MCP tool outputs. While this constitutes an indirect prompt injection surface, the skill mitigates risk through the following measures:
Ingestion points: Evaluation questions are parsed from XML via parse_evaluation_file; tool outputs are received through the MCP connection in the agent_loop function.
Boundary markers: The EVALUATION_PROMPT enforces a strict output format using XML tags (<summary>, <feedback>, <response>) to separate the model's internal processing from its final output.
Capability inventory: The script can execute local subprocesses (to launch servers) and make network requests to the Anthropic API and external MCP endpoints.
Sanitization: All tool outputs are stringified or JSON-encoded before being incorporated into the prompt history to prevent structure-based confusion.

mcp-builder