Guides

Claude Code and Codex in one app

Claude Code, from Anthropic, and Codex, from OpenAI, are both coding agents that run in a terminal: they read your project, edit files and run commands. Many developers end up with both installed. This guide covers why using both can help, how to share one project between them, how to run them on your own plans, and how Yami puts them in one app.

Why use both

• A second opinion from another model family. Different models tend to miss different things. Asking the other agent to review a change, or to explain a bug the first one is stuck on, is a cheap check.
• Two allowances. If you have both a Claude plan and a ChatGPT plan, the other agent can carry on when one plan's usage limit is reached in the middle of a task.
• Fit for the task. How well an agent does depends on the language, the codebase and the kind of task, and the models change often. The practical way to know for your own code is to give both the same task and compare the results.

What each one is, in plain terms

Claude Code runs Anthropic's Claude models. You sign in with a Claude plan or use an Anthropic API key. It reads project instructions from a CLAUDE.md file, asks before edits and commands according to the permission mode you choose, and supports hooks, subagents and MCP servers for extra tools. It keeps its sessions under ~/.claude, and claude --continue or claude --resume picks one up again.

Codex runs OpenAI's models. You sign in with a ChatGPT plan or use an OpenAI API key. It reads project instructions from an AGENTS.md file and runs commands inside an operating system sandbox whose limits you set, such as writing only inside the project folder. codex exec runs it without the interactive screen, for scripts. It keeps its sessions under ~/.codex, and codex resume picks one up again. It supports MCP servers too.

Both are under active development, so check each one's --help for the options your version has.

Sharing one project between them

Keep one instructions file. Claude Code reads CLAUDE.md and Codex reads AGENTS.md. Rather than keeping two copies in step, put your instructions in AGENTS.md and make CLAUDE.md a single line that imports it; Claude Code's memory files read a line starting with @ as an import:

echo "@AGENTS.md" > CLAUDE.md

Commit between agents. Before you hand a task from one agent to the other, commit. Each agent's work then sits in its own commits, easy to compare or revert:

git add -A
git commit -m "Login form, first pass"
git log --oneline -5

Do not run both in one folder at the same time. Two agents in one checkout edit the same files and pick up each other's half finished changes. If you want them working side by side, give each its own git worktree and branch; the guide to git worktrees for AI coding agents shows how.

Hand over context in writing. The two agents do not read each other's sessions. Ask the first one to write a short summary of what it did and what is left, in a notes file or the commit message, and start the second one from that.

Use the official CLIs with your own plans

Both plans are meant to be used through each company's own tools. The plain way to stay inside their terms is to install Claude Code and Codex from their makers' documentation, sign in yourself, and run them as they ship:

claude        # the first run asks you to sign in
codex login

That also keeps each tool's own permission checks, sandbox and updates working as their makers intended. Be careful with tools that copy a plan's sign-in token into another program to call the models directly: that is a different thing from running the official CLI, and may not be allowed by your plan. Also keep in mind that a personal plan is made for one person; many agents running in parallel or on a schedule may not fit its terms, and an API key billed per token is made for that use.

The catch: two histories

With two agents, the story of one feature can end up split across two places: some of it in Claude Code's sessions and some in Codex's, each with its own resume command. Notes in the repository and clear commit messages help. So does an app that shows both in one list.

How Yami does it

Yami is a free desktop app for macOS (Apple silicon) and Windows 11 that works with Claude Code and Codex. It does not bundle or replace them. If you have them installed and signed in, Yami runs those programs on your own plans, with their own settings, and you pick the engine for each chat.

• One list of chats. Chats run on Claude Code and on Codex sit side by side in Yami, by project.
• Your earlier sessions too. Import chats reads your existing Claude Code and Codex sessions (and OpenCode, Gemini CLI, Cline, Roo Code, Aider, and ChatGPT or Claude.ai exports) without changing them. An imported Claude Code or Codex chat carries on where it left off, as long as its folder still exists.
• Approvals in one place. Claude Code's permission questions appear as Yami's approval cards. Codex keeps its own sandbox.
• Other engines when you want them. The same app also runs a built-in engine (DeepSeek Harness, whose provider's servers are in China), API keys (Anthropic, OpenAI-compatible or OpenRouter) and local models.
• On your computer. Chats, files, memory and keys stay on your computer, and Yami itself has no analytics. Each engine still sends what it needs to its own model, as it would without Yami.

For bigger jobs, a venture lets a leader agent plan the work and a crew of agents build the parts in separate worktrees; the guide on running several agents in parallel explains how.

Claude and Claude Code are trademarks of Anthropic, PBC. Codex and ChatGPT are trademarks of OpenAI. Yami works with both tools but is independent: it is not made by, affiliated with or endorsed by Anthropic or OpenAI.