If you need to run multiple Codex instances with different personalities and permissions, this is the wrapper for you. It lets you spin up specialized agents from a single binary, each configured independently via environment variables. You might run a strict code reviewer locked to read-only sandbox, a test writer with workspace-write access, and a quick helper on minimal effort, all appearing as separate tools in your MCP client. It wraps the OpenAI Codex SDK and exposes a task tool plus a reply tool for threading conversations. There's also a factory mode that generates configuration snippets interactively. Useful when one general-purpose Codex isn't enough and you want purpose-built agents without managing separate installations.
claude mcp add --transport stdio xiaolai-codex-octopus uvx codex-octopus