The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/connections.py and scripts/evaluation.py files contain logic to execute shell commands provided via CLI arguments. This is the intended mechanism for the evaluation harness to launch and interact with a local MCP server process under test via the standard input/output (stdio) transport.
[EXTERNAL_DOWNLOADS]: The skill instructions in SKILL.md direct the agent to fetch documentation from the official modelcontextprotocol.io website and official SDK README files from the modelcontextprotocol GitHub organization. These are recognized as trusted and well-known sources for this protocol.
[INDIRECT_PROMPT_INJECTION]: The evaluation harness in scripts/evaluation.py ingests and processes data returned by MCP server tools to generate reports. This presents a surface for indirect prompt injection if a tool returns malicious instructions intended to influence the evaluation agent.
Ingestion points: Data enters the agent context through connection.call_tool results in scripts/evaluation.py.
Boundary markers: The EVALUATION_PROMPT uses XML tags (<summary>, <feedback>, <response>) to delimit agent output, providing structure that helps mitigate accidental instruction obedience.
Capability inventory: The evaluation script has the capability to call any tool registered by the connected MCP server and communicate with the Anthropic API.
Sanitization: Tool outputs are converted to strings without specific sanitization before being included in the prompt history for the LLM.
[CREDENTIALS_UNSAFE]: The documentation in reference/evaluation.md and scripts/evaluation.py correctly instructs users to manage sensitive credentials like ANTHROPIC_API_KEY or GITHUB_TOKEN using environment variables. This follows security best practices for credential management in development tools.

mcp-builder