codex-mcp-server

The server implements robust input validation using Zod and utilizes 'cross-spawn' for external command execution, which is generally safer than raw shell execution for arguments. It transparently exposes the underlying Codex CLI's sandbox modes ('read-only', 'workspace-write', 'danger-full-access') and approval policies, explicitly flagging the 'yolo' (bypass all safety) option as dangerous. No hardcoded secrets were found; API key management is delegated to the Codex CLI (via 'codex login' or 'OPENAI_API_KEY' environment variable). Overall safety heavily relies on the user's careful configuration and understanding of the Codex CLI's sandbox and approval settings, as the server delegates powerful operations to the AI agent.