codex-mcp-go

The server acts as a wrapper for the `codex` CLI, passing user-provided prompts as arguments. It incorporates sandbox controls (`read-only`, `workspace-write`, `danger-full-access`) for the underlying `codex` operations. A significant security consideration is the default `yolo=true` flag, which causes the `codex` CLI to execute generated commands without explicit human confirmation. This design, while intended for automated agentic workflows, introduces a risk where a malicious or erroneous AI prompt could lead to unintended actions within the configured sandbox. No direct `eval` or similar dangerous patterns are used within the Go code, nor are there hardcoded sensitive secrets. Path validation for `cd` and `image` parameters is present. The overall security relies heavily on the robustness of the underlying `codex` CLI and the trust in the AI's generated actions when `yolo` is enabled.