Workflow · 6 November 2026 · 7 min read

Claude Code and Codex together: a daily workflow

Using Claude Code and Codex together is less about picking a winner than about having two agents and two plans at hand. Claude Code comes with a Claude plan, Codex with a ChatGPT plan, and both run in the same terminal, on the same repository. This guide shows a setup that keeps them from stepping on each other: one instructions file, a split of tasks that plays to each one, and a single place to see what both are doing.

In short

Put your project rules in AGENTS.md: Codex reads it, and recent Claude Code versions read it when the project has no CLAUDE.md (or import it with @AGENTS.md).
Give each agent its own task, or its own worktree, never the same files at the same time.
Use the second agent as a reviewer: a fresh model on a diff catches what the first one missed.
When one plan hits its limit, the other is untouched: the limits are separate.
Tarmac shows the sessions and the plan usage of both agents for every project, with one task list they both keep up to date.

Why developers keep both

Three reasons come back in most write-ups and forum threads:

Two plans, two sets of limits. A Claude plan and a ChatGPT plan have separate 5-hour and weekly windows. When one runs out mid-afternoon, the work goes on with the other.
A second opinion. Two different models make different mistakes. Asking the agent that did not write the code to review it is cheap and often useful.
Different strengths on different days. Opinions on which agent is better at what change with every model release. Having both installed means you can try the same task with each instead of guessing.

1. One instructions file for both

Each agent reads a file of project instructions at the start of every session: Codex reads AGENTS.md, Claude Code reads CLAUDE.md. Keeping two copies in sync is the first thing that breaks. Two ways to keep one:

Only AGENTS.md. Since version 2.1.277, Claude Code reads AGENTS.md as your project instructions when there is no CLAUDE.md in the folder or above it. Nothing else to set up.

A CLAUDE.md that imports it. If you need Claude-specific rules too, keep the shared ones in AGENTS.md and import it:

@AGENTS.md

## Claude Code only

Use plan mode for changes under `src/billing/`.

Write in that file what both agents must respect: how to run the tests, which commands are off limits, the branch rules.

2. Split the work, not the files

The rule from running several Claude Code sessions applies here too: two agents in the same folder share the same files. Three patterns that work:

Different tasks, different areas. Claude Code on the API, Codex on the front end. Say so in the prompt and name the folders.
Different checkouts. For two changes in the same area, give each agent its own git worktree. Claude Code creates one with claude --worktree <name>; for Codex, create it yourself and start codex inside it:
```
git worktree add ../shop-codex -b codex/filters
cd ../shop-codex && codex
```
Writer and reviewer. One agent writes, the other only reads. This is the safest way to start.

3. The review loop

Once a change is ready, open the other agent in the same folder and ask for a review of the diff, for example:

Review the uncommitted changes in this repository. List bugs, missing tests and
anything that breaks the rules in AGENTS.md. Do not edit files.

Codex also has a /review command for this. Then take the findings back to the first agent, which still has the context of why the code looks the way it does. “Do not edit files” matters: a reviewer that starts fixing things is a second writer on the same files.

4. Switch when a limit runs out

Both agents show their limits on demand: /usage in Claude Code, /status in Codex. When one says the 5-hour window is full, finish the task with the other. What makes the hand-over smooth:

The plan is written down. A task list in the repository, or a plan file, tells the second agent where the first one stopped.
The instructions are shared (step 1), so the second agent follows the same rules.
Commits are small. The second agent reads the history to see what was done.

The details of each plan’s limits are in how to check your Claude Code usage limit, which covers Codex too.

5. See both in one place

This is where the setup gets clumsy without help: two agents, several projects, sessions in a dozen terminals, two usage screens. Tarmac, a free and open source local dashboard for macOS and Windows (WSL 2), puts both agents next to the projects they work on:

A Claude tab and a Codex tab in every project, with every session, active or finished, including those started in a terminal or an IDE.
New or resumed sessions of either agent in a console built into Tarmac (codex resume <id> for Codex).
Both plans in the sidebar: Claude’s 5-hour, weekly and per-model limits, and ChatGPT’s 5-hour and weekly windows, live.
One task list per project that both agents keep up to date through Tarmac’s MCP server (/suivi in Claude Code, $suivi in Codex), with a Todolist/ copy in the repository.
A ▶ button that sends the next pending task to an idle Claude or Codex session running in Tarmac, or to a new one.

To turn Codex on, tick it in Settings → Agents and run tarmac setup --with-codex. Claude Code alone is the default; Codex alone works too. The full tour is on the Claude Code and Codex page.

A starting checklist

Both agents installed and signed in with their plan (Claude Code with /login, Codex with your ChatGPT account).
One AGENTS.md at the root of each project, imported from CLAUDE.md if you keep one.
One rule: never two agents writing the same files at the same time.
A review prompt saved somewhere you can paste it.
One view of all sessions and both plans.

Next steps

Install Tarmac, tick Codex in Settings, and open a project: its Claude and Codex tabs show what each agent has done there.

Sources: Claude Code documentation, How Claude remembers your project (AGENTS.md support) and Worktrees; OpenAI, Codex commands and Codex pricing. Checked on 6 November 2026.