The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to fetch documentation and SDK README files from the official Model Context Protocol website (modelcontextprotocol.io) and its associated GitHub repository. These are well-known and trusted sources for the protocol being documented.
[COMMAND_EXECUTION]: The provided scripts/evaluation.py script is designed to execute user-provided shell commands (via the -c and -a flags) to launch local MCP servers for automated testing. This is a standard and necessary function for a development/testing harness.
[REMOTE_CODE_EXECUTION]: While the evaluation script can execute a local server process, it does so based on explicit user input for the purpose of testing a development project. There is no evidence of unauthorized or hidden remote code execution patterns.
[DATA_EXFILTRATION]: The skill follows security best practices by recommending the use of environment variables (e.g., ANTHROPIC_API_KEY) for sensitive credentials rather than hardcoding them. Network operations are limited to communicating with the specified MCP server endpoints and the Anthropic API.
[INDIRECT_PROMPT_INJECTION]: The evaluation.py script implements an agent loop that passes outputs from MCP server tools directly to the LLM. This represents a potential surface for indirect prompt injection if the server returns malicious content, though this is an inherent risk in testing external integrations.
Ingestion points: scripts/evaluation.py (receives tool result output via connection.call_tool).
Boundary markers: None implemented for tool results passed into the message history for the model.
Capability inventory: The script can invoke any tool exposed by the connected MCP server and interacts with the Anthropic API to process these results.
Sanitization: Tool outputs are converted to strings or JSON for the message block but are not sanitized for malicious instruction patterns.

mcp-builder