The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/evaluation.py and scripts/connections.py files utilize the mcp library to launch local MCP servers. This involves executing shell commands provided as arguments (e.g., via the -c and -a flags) to start the server process using standard input/output streams.
[INDIRECT_PROMPT_INJECTION]: The evaluation harness in scripts/evaluation.py processes responses from external MCP servers. These responses are incorporated into the agent's context to facilitate multi-step task completion and evaluation. If an MCP server returns malicious instructions, it could potentially influence the evaluator agent's behavior.
Ingestion points: The agent_loop function in scripts/evaluation.py receives data from the call_tool method which communicates with external servers.
Boundary markers: The skill relies on standard tool_result message structures but does not implement additional sanitization or explicit safety boundary markers for the returned content.
Capability inventory: The evaluator agent has the ability to call any tool exposed by the MCP server being tested, which may include network or file system access.
Sanitization: None observed; tool outputs are treated as data to be evaluated and may be directly appended to the conversational history.
[EXTERNAL_DOWNLOADS]: The skill instructions and implementation guides fetch documentation and SDK information from official Model Context Protocol repositories on GitHub (modelcontextprotocol organization).

mcp-builder