How to Run Claude Code and Codex Together on One Project (2026)
Most teams now have more than one coding agent in play. One engineer lives in Claude Code, another in Codex, someone else in Cursor. Each agent is good on its own. The trouble starts when they work on the same codebase: they don’t know about each other, so the engineers end up relaying plans, warnings and findings between them by hand.
There are four ways to fix that, and they solve different problems. Which one fits depends on a single question: are you coordinating one person’s agents, or several people’s?
The short answer
- One person, one vendor: Claude Code’s agent teams.
- One person, several agents in parallel: a parallel-agent workspace such as Paperclip, Conductor or Nimbalyst, with a git worktree per task.
- Several people, each with their own agent: a shared thread where the agents coordinate with each other, such as elvex.
- Anything, with no new tools: git worktrees, pull requests and a shared AGENTS.md.
1. Git worktrees and pull requests
The no-new-tools option. Give each agent its own git worktree and branch, so two agents never edit the same checkout, and make the pull request the hand-off point. Put the shared rules in an AGENTS.md file at the root of the repo. Codex reads it directly, and Claude Code can be pointed at it from CLAUDE.md.
It works, and it keeps a clean history. What it doesn’t do is let the agents talk. If Codex is about to rename a field that Claude Code still reads, nobody finds out until the merge.
2. Claude Code agent teams
Claude Code can coordinate several of its own sessions as a team. One session acts as the lead and assigns tasks, and the teammates share a task list and message each other directly. It’s the most integrated option if everything runs in Claude Code.
The limits are in the name. It coordinates Claude Code sessions, not Codex or Cursor, and the sessions belong to one person. Anthropic also marks agent teams as experimental and off by default, with known limits around resuming sessions and shutting down.
3. Parallel-agent workspaces
A growing set of tools runs many coding agents side by side, each task in its own worktree, with a person reviewing before anything merges. In their own words:
- Paperclip runs Claude Code, Codex, Gemini CLI, Cursor and others on your repos, gives each task its own git worktree and one owner, and merges nothing until a person approves. It’s open source.
- Conductor runs Claude Code, Codex, Cursor and OpenCode in cloud sandboxes and has added multiplayer workspaces you can share with teammates. It has announced that it is joining Meta.
- Nimbalyst is a visual workspace for Codex and Claude Code: sessions, tasks, docs and diffs in one place, using your existing subscription.
- Vibe Kanban puts agent tasks on a kanban board and works with Claude Code, Codex and OpenCode. The company has announced it is winding down, and the project continues as community-maintained open source.
- Coordigent plugs into Claude Code, Codex, Cursor, Windsurf, Gemini CLI and others so each agent can see what the others are working on.
These are strong at running many agents at once and keeping their changes apart. Most of them start from one person’s machine and one person’s queue of tasks.
4. Shared threads across people and harnesses
The fourth approach starts from the team instead of the task queue. Each person keeps the agent they already use, with its skills, memory and permissions, and connects it to a shared thread. The agents read the same thread, claim and hand off to-dos, flag conflicts before they land, and bring a person in only when a decision is needed.
That is what elvex does. Diego’s Codex can see that Priya’s Claude Code still reads a field in six places before it renames it, and ask Priya which way to go. Claude Code can write failing tests and hand the fix to Codex, which has the billing repo. The agents connect over MCP, so nobody switches harness, and inference stays on each person’s own subscription.
How they compare
| Approach | Agents | Across people? | How agents coordinate |
|---|---|---|---|
| Worktrees + PRs | Any | Yes, through Git | They don’t; people do |
| Claude Code agent teams | Claude Code only | No | Shared task list and messages |
| Parallel-agent workspaces | Claude Code, Codex and more | Some, via shared workspaces | Task board, worktree per task, review |
| Shared threads (elvex) | Claude Code, Codex, Cursor, any MCP client | Yes | Same thread, shared to-dos, questions to people |
You can combine them. Worktrees and pull requests stay useful underneath any of the other three.
Frequently Asked Questions
Can Claude Code and Codex work on the same repository at the same time?
Yes, as long as they don’t share a working directory. Give each agent its own git worktree and branch, and merge through pull requests. To stop them from colliding on the same code, they also need a way to see each other’s plans, which is what agent teams, parallel-agent workspaces and shared threads add.
Do Claude Code and Codex read the same instructions file?
Not by default. Codex reads AGENTS.md, and Claude Code reads CLAUDE.md. A common fix is to keep the shared rules in AGENTS.md and reference it from CLAUDE.md.
What is the difference between multi-agent and multiplayer?
Multi-agent means several agents working on one task, often for one person. Multiplayer AI means several people and their agents working together, so the coordination between people moves to the agents.
Does elvex replace Claude Code or Codex?
No. elvex connects to the agents you already use over MCP. They keep running where they are, on your own subscription, and coordinate through elvex.