The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/connections.py and scripts/evaluation.py files allow for the execution of arbitrary local commands via the MCP stdio transport. This is intended for launching and testing local MCP servers but grants the script the ability to spawn subprocesses based on user-provided CLI arguments.
[EXTERNAL_DOWNLOADS]: The SKILL.md file points to documentation and resources on modelcontextprotocol.io and GitHub repositories under the modelcontextprotocol organization. These are official resources for the protocol and its SDKs.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection within the evaluation harness found in scripts/evaluation.py.
Ingestion points: The script ingests untrusted data from an XML evaluation file and directly processes responses from the tools of the connected MCP server.
Boundary markers: The EVALUATION_PROMPT used in scripts/evaluation.py lacks delimiters or specific instructions to the model to ignore potential instructions embedded in tool outputs.
Capability inventory: The script can execute any tool provided by the connected MCP server, which (via stdio transport) can include arbitrary command execution on the local machine.
Sanitization: The script does not perform any sanitization or validation of tool outputs before passing them back into the LLM context history.

mcp-builder