Claude Code and Codex together: one project, two agents, no collisions
Claude Code and Codex are good at different things, and a second agent that reads the first one's work catches mistakes neither would catch alone. The hard part is not running both. It is keeping them from editing the same files at the same moment and keeping track of who did what.
Below are three setups, from plain terminals to SHIPIT, with the trade-offs of each.
Why use both
- A second opinion. One agent writes the change, the other reviews the diff with fresh context.
- Different strengths. Teams often keep one model for long refactors and the other for quick fixes and tests. Pick by results on your own code, not by reputation.
- Two plans, two sets of limits. When one subscription window runs low, the other agent keeps going.
Setup 1: two terminals in one folder
Open a terminal for claude and another for codex in the same repository. It is the quickest start and works for review: one agent changes code, the other only reads. If both edit at once they overwrite each other, so give the reviewer read-only instructions.
Setup 2: one Git worktree per agent
Create a worktree for each agent so each works on its own branch in its own folder:
git worktree add ../app-codex -b codex/task
Run Claude Code in the main folder and Codex in ../app-codex, then merge the branch you like. No collisions, at the cost of managing the folders and merges yourself.
Setup 3: two teammates in SHIPIT
SHIPIT runs both in one Mac window and keeps the conversation between them visible:
- Connect both engines in Settings, Agents & AI. SHIPIT installs Claude Code and Codex if they are missing and opens the sign-in for your existing plans.
- Create two teammates, for example Builder and Reviewer. Each chat picks its model from the installed engines, so Builder can run on a Claude model and Reviewer on a Codex model.
- Give Builder the task. When it is done, type
@Reviewer check the diff. The handoff card shows the state (queued, working, needs you, replied) and the review comes back to your chat. - Keep Ask before acting on: every file write and tool run waits for your Approve.
- Need them on separate branches? Use Create Git worktree from the pane menu and open the second agent there.
The Dashboard shows how much of each Claude and Codex subscription window is left, so you can see which agent should take the next task.
Which setup to choose
| Setup | Good for | Watch out for |
|---|---|---|
| Two terminals | Quick review by a second agent | Both editing at once |
| Worktrees | Two agents changing code in parallel | Folders and merges are on you |
| SHIPIT teammates | Handoffs, approvals and limits in one place | Mac only, Dzhura subscription |
Questions and answers
Can Claude Code and Codex talk to each other directly?
Not by themselves. Something has to pass the work along: you copying text between terminals, a shared file, or a tool like SHIPIT where @Name hands a task from one teammate to another and the reply comes back.
Do I need two subscriptions?
Yes. Claude Code runs on an Anthropic plan and Codex on an OpenAI plan. Each agent spends its own limits.
Will two agents break my repository?
Only if they edit the same files at the same time. Keep one of them read-only, or give each its own Git worktree.