The Agent Skills Directory

[COMMAND_EXECUTION]: The ask_codex tool is a wrapper for the Codex CLI that allows the agent to execute arbitrary tasks on the local system. The documentation explicitly lists 'Install dependencies and run tests' as examples of use.
[PRIVILEGE_ESCALATION]: The skill implements a sandbox_mode parameter for the ask_codex tool. The danger-full-access mode provides the agent with full system access, which bypasses standard workspace boundaries and safety constraints.
[DATA_EXFILTRATION]: The ask_many and review tools facilitate sending potentially sensitive project content, file contents, or code snippets to multiple third-party LLM providers (OpenAI, DeepSeek, Moonshot, Alibaba) in parallel. This creates a large data exposure surface if the agent is prompted to analyze sensitive files.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of external Python packages (fastmcp, openai) and assumes the presence of a pre-installed and authenticated codex CLI tool.
[INDIRECT_PROMPT_INJECTION]: Because this skill is designed to analyze codebase structure and content (e.g., 'Find all TODO comments', 'Explain the architecture'), it is highly susceptible to indirect prompt injection if the project being analyzed contains malicious instructions hidden in comments or documentation meant to influence the orchestrating LLM or the Codex execution.

multi-llm-mcp-server