The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches protocol specifications and SDK documentation from official sources, including modelcontextprotocol.io and the modelcontextprotocol organization on GitHub. These references are documented neutrally as they originate from well-known service providers for the Model Context Protocol ecosystem.
[INDIRECT_PROMPT_INJECTION]: The skill includes an evaluation harness that processes outputs from external MCP servers and reads documentation from the web, which constitutes a potential attack surface for indirect prompt injection.
Ingestion points: External API documentation fetched via WebFetch and tool execution results processed by the evaluation script in scripts/evaluation.py.
Boundary markers: The system prompt in the evaluation script uses XML delimiters to structure agent responses but does not include specific instructions to ignore commands potentially embedded within external data.
Capability inventory: The evaluation script is capable of invoking tools on connected MCP servers, which could grant access to file systems or network operations depending on the specific server's configuration.
Sanitization: External data content and tool outputs are processed without explicit sanitization or filtering before being passed back to the model context.

mcp-builder