mcp-builder
Pass
Audited by Gen Agent Trust Hub on Jul 23, 2026
Risk Level: SAFECOMMAND_EXECUTION
Full Analysis
- [SAFE]: Fetches Model Context Protocol specification and SDK documentation from official sources, including the modelcontextprotocol.io domain and the modelcontextprotocol GitHub organization.
- [COMMAND_EXECUTION]: The evaluation harness in scripts/evaluation.py and connection utility in scripts/connections.py use the mcp library to spawn subprocesses for the stdio transport. This functionality is intended to launch and communicate with local MCP servers for development and testing purposes.
- [SAFE]: The toolkit includes a specialized system prompt for an evaluation agent in scripts/evaluation.py. While the processing of tool outputs from external servers represents an indirect prompt injection surface (Category 8), the implementation includes boundary markers and is intended for controlled development use.
- Ingestion points: Tool output results from MCP servers processed in the agent loop within scripts/evaluation.py.
- Boundary markers: The EVALUATION_PROMPT explicitly instructs the agent to wrap output segments in , , and XML tags.
- Capability inventory: The agent can invoke any tool exposed by the connected MCP server using the call_tool method.
- Sanitization: No explicit content sanitization or escaping is performed on the results returned by the MCP tools before interpolation into the conversation history.
Audit Metadata