The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides a connection utility (scripts/connections.py) and an evaluation harness (scripts/evaluation.py) that use the mcp SDK to execute local shell commands. This capability is used to start and manage MCP server processes using the standard I/O (stdio) transport as part of its primary testing functionality.
[INDIRECT_PROMPT_INJECTION]: The evaluation harness in scripts/evaluation.py processes external data that could contain malicious instructions.
Ingestion points: The script reads test questions from XML files and ingests raw output from MCP tools during the agent loop.
Boundary markers: The agent is constrained by a system prompt (EVALUATION_PROMPT) requiring it to use specific XML tags (<summary>, <feedback>, <response>) for its output.
Capability inventory: The skill can execute local subprocesses to run servers and initiate network connections via SSE or HTTP transports to interact with MCP endpoints.
Sanitization: External tool results are stringified or JSON-encoded before being re-interpolated into the LLM context to prevent direct instruction leakage.

mcp-builder