How to use Claude & Codex together

Most AI coding conversations are framed like a competition. Claude or Codex? Anthropic or OpenAI? Paid or open-source models? American or Chinese ones? This model or that model. Which one writes better code? Which one reasons better? Which one should your team standardize on?

Argh.

That framing is understandable, but it misses the more useful pattern.

For serious development work, Claude, Codex and other models should not be treated as mutually exclusive tools. They should be used TOGETHER.

Why can't one model review its own code?

It can, but it reviews through the same reasoning it used to write the code. It knows what it meant, so it reads the intent instead of the text, and it tends to miss the same edge case twice.

People work the same way. A developer who reviews their own pull request rarely finds every bug, because they check it with the same assumptions they wrote it with. That is why teams have code review and QA: not because developers are careless, but because self-review has limits.

Research on language models points the same way. One study found that models struggle to correct their own reasoning without outside feedback, and sometimes make answers worse when they try (Huang et al., "Large Language Models Cannot Self-Correct Reasoning Yet"). Another that models used as evaluators recognize and rate their own outputs higher than others' (Panickssery et al., "LLM Evaluators Recognize and Favor Their Own Generations"). A second model from a different vendor has neither the same blind spots nor a reason to like the first model's work.

How to run Claude and Codex together?

Sure, you could copy and paste things from window to window. But that way you work on one task at a time and turn into a copy-paste machine. That is not the future we deserve.

Another option is to run Codex inside a Claude Code session through OpenAI’s plugin. This shortens the copy and paste, but the task, the review and the decisions still disappear when the session ends. And actually this doesn’t work if you want to use Codex or other provider as main coding agent.

The way to go is an agentic task board with a built-in coding harness, like Morion. It is a desktop app that keeps everything in one place: tasks your agents work from over MCP, a memory layer that hands each ticket the notes it needs, and Auto-code, which runs a ticket through one or several agents. There are two ways to use it for this task.

Agent-managed. Let’s say you picked Claude as your main coding agent. Ask it to create a ticket in Morion and add a rule: after every iteration, the ticket goes through a review-only Auto-code workflow where Codex reviews the change. You talk only to Claude, and the Codex reviews happen on their own. You see every iteration, the review comments and the changes on the ticket, and the ticket keeps the whole history, as if a colleague had worked on it. Auto-code has no review-only template out of the box: build it once in the workflow editor, or ask Mo to build it for you.

Fully automated. Here you set up Auto-code as a full workflow with several agents. You add a task (or ask an agent to add it) and move it to Todo in a folder. That’s it: a workflow runs it. The default template, “Code + code review · Mo-driven”, has Claude write the change in its own git worktree, Codex review it, and Mo decide whether it is done or goes back.

Add a human_gate step and the run pauses to ask you in chat when the agents cannot settle an issue. If the work fails, the ticket returns to Re-opened with the reason, and a coding agent can pick it up to fix the issues. Any slot can run claude, codex, cursor, opencode or pi, so the coding agent or the reviewer can be an open model too.

All Your notes and tasks stay only on your machine.

Try it on your next ticket

Download Morion for macOS, Windows or Linux. New accounts get 14 days of Pro, no card.

Sources: openai/codex-plugin-cc README (checked 2026-09-26); Huang et al., 2023; Panickssery et al., 2024.