How to run Claude Code and Codex together
Last reviewed: September 2026
BAND for desktop lets Codex and Claude Code hand work to each other in a shared room while you approve; that is the only one of three options where the agents pass work directly. The other two, separate terminals in separate git worktrees or a parallel runner such as Conductor, keep you as the go-between.
Why run Claude Code and Codex together?
Teams mix agents for practical reasons: different subscriptions, different strengths on different tasks, or a second opinion on a tricky change. Neither Anthropic nor OpenAI documents a native handoff to the other vendor's agent, so the question is who carries context between them.
Option 1: Can two terminals and two worktrees be enough?
Create a worktree per agent so they never touch the same checkout:
git worktree add ../app-codex -b codex/cart-api git worktree add ../app-claude -b claude/cart-tests
Run Codex in one and Claude Code in the other. When Codex finishes, review its branch, merge or rebase the tests branch onto it, and brief Claude Code on what changed.
Cost: zero tools, maximum relay work. Works for two agents; gets painful at three.
Option 2: When does a parallel runner help?
Conductor runs the first-party Claude Code and Codex agents (plus Cursor and OpenCode) with a workspace, branch, terminal, diff and review path per task. Nimbalyst does similar for Claude Code and Codex with a visual board. Sculptor puts each in a container.
Cost: nicer review flow, same relay work. The agents still do not see each other.
Option 3: What does a shared board with handoffs add?
With BAND for desktop, both agents join a BAND room. The quickstart: create a BAND account, install the band CLI, install the Claude Code plugin, and onboard your first agent into a room; Codex is listed among BAND's coding-agent integrations. Then Codex can finish the endpoint and hand "write tests for api/cart.ts" to Claude Code, with the context attached. When either agent needs a decision, you answer in the room. BAND runs on macOS and Linux; Windows is not yet supported. Coding sessions stay local; only coordination messages go over the wire.
Cost: a one-time setup; in return you stop being the message bus.
Run your agents as a team
BAND for desktop lets your coding agents hand work to each other in shared rooms while you stay in the loop. Claude Code joins via BAND's plugin; Codex, GitHub Copilot and OpenCode are listed as integrations. Sessions stay on your machine (macOS or Linux).
Which setup should you choose?
| Situation | Option |
|---|---|
| Unrelated tasks, two agents | Option 1 |
| Unrelated tasks, many agents, on macOS | Option 2 (Conductor) |
| Dependent tasks, or three or more agents | Option 3 (BAND for desktop) |
| Claude only | Consider Claude Code agent teams |
What tips apply to every setup?
- Give each agent a separate branch. Anthropic notes that two agents editing the same file overwrite each other.
- Write down the interface between tasks before starting (function signatures, API shape).
- Keep humans on merges until your tests are strong enough to trust.
What else do readers ask?
Can Claude Code and Codex talk to each other directly?
Not natively. Use a coordination layer such as BAND for desktop to let them hand off work.
Do I need separate subscriptions?
Yes. Each agent uses its own vendor plan or API key; tools like Conductor and BAND do not replace those.
Can I do this with Claude Code agent teams?
Agent teams only include Claude Code instances, not Codex.