Skip to content
HomeDocsPricingDownload
Mixed-provider agents

How to run Claude Code, Codex and Gemini CLI together on one project

Install and log in to each CLI, then give each agent its own terminal and, when they edit code at once, its own git worktree. That handles the files; hand-offs between agents stay manual. MAO runs all three as terminal nodes on one canvas, each on its own login, with wires for hand-offs and a role per agent.

  • Each CLI keeps its own login and plan
  • One git worktree per agent that edits code
  • MAO needs Gemini CLI 0.41 or later
By Updated 11 min read

To run Claude Code, Codex and Gemini CLI on one project, install and sign in to each CLI, start each in its own terminal, and give every agent that edits code its own git worktree so their changes stay apart. Plain terminals and worktrees leave you to pass plans and reviews between the agents. MAO runs all three as terminal nodes on one canvas, each on its own login, with wires that let them send each other tasks and a role for each.

Why mix Claude Code, Codex and Gemini CLI?

Each CLI is a different company's agent for its own models, so mixing them lets you give each part of the work to a different model. A common split is one agent to plan, one to build and one to review.

  • Claude Code is Anthropic's agent and signs in with a Claude subscription or a Claude Console account.
  • Codex is OpenAI's agent and signs in with a ChatGPT plan or an OpenAI API key.
  • Gemini CLI is Google's agent and signs in with a Google account, a Gemini API key or Vertex AI.

A reviewer that did not write the code reads it fresh, and one from another provider is a further step removed from the author. Which model suits which job depends on your project, so try a split on a small task before you rely on it.

What does each CLI need before you start?

Each one installs separately and asks you to sign in the first time you run it in a project directory.

Claude CodeCodexGemini CLI
InstallNative installer, Homebrew or WinGetInstall script, npm or Homebrewnpm, Homebrew or MacPorts; needs Node.js 20 or later
Startclaudecodexgemini
Sign inClaude Pro, Max, Team or Enterprise, Claude Console, or a cloud providerSign in with ChatGPT (Plus, Pro, Business, Edu, Enterprise) or an API keySign in with Google, a Gemini API key, or Vertex AI
Project instructionsCLAUDE.md, or AGENTS.mdAGENTS.mdGEMINI.md

With Homebrew:

brew install --cask claude-code
brew install --cask codex
brew install gemini-cli

On Windows:

winget install Anthropic.ClaudeCode
npm install -g @openai/codex
npm install -g @google/gemini-cli

Claude Code and Codex also have their own install scripts for each system, listed in their docs. Claude Code's native install updates itself in the background; a Homebrew or WinGet install does not, so upgrade it yourself with brew upgrade claude-code or winget upgrade Anthropic.ClaudeCode.

Then sign in to each one from your project directory:

  1. Run claude --version to check the install, then claude and follow the browser login. Type /login inside a session to switch accounts later.
  2. Run codex and choose Sign in with ChatGPT, or sign in with an API key by running codex login --with-api-key with the key on its standard input. codex login status shows which method is active.
  3. Run gemini and choose a sign-in method. For a Gemini API key, set it first with export GEMINI_API_KEY="YOUR_API_KEY". With a Code Assist licence from your organization, set GOOGLE_CLOUD_PROJECT to your Google Cloud project before you sign in with Google.

Check which Google accounts Gemini CLI serves before you set it up. Its README says it is meant for Gemini Code Assist Standard and Enterprise licences and Code Assist for GitHub on Google Cloud, plus paid API and Vertex AI keys. Free individual access and Google AI Pro and Ultra moved to Antigravity CLI on June 18, 2026.

How do the three agents share project instructions?

Keep one AGENTS.md and point the other two at it. Each agent reads its own instruction file by default, so without this you maintain three copies of the same rules.

  • Codex reads AGENTS.md, walking from the project root down to the current directory.
  • Claude Code reads AGENTS.md when there is no CLAUDE.md. If you keep a CLAUDE.md, add a line @AGENTS.md to it to import the shared file.
  • Gemini CLI reads GEMINI.md. List other names in the context.fileName setting of its settings.json:
{
  "context": {
    "fileName": ["AGENTS.md", "GEMINI.md"]
  }
}

Can you just open three terminals?

Yes, and it needs nothing new: open three terminal windows or tabs in the repository and run claude, codex and gemini, one in each. Each CLI keeps its own login, so the three never share credentials.

It works while the agents read code or touch clearly separate files. They all edit the same working tree, though, so one agent's half-finished change is what the next one sees, and you are the only link between them: you copy the plan out of one window and paste it into the next.

Does tmux make it easier?

tmux puts the three agents in panes of one window, so you can watch them side by side and leave them running in a session you come back to. It changes how you see the agents, not how they share files.

  1. Start a named session with tmux new -s agents.
  2. Press C-b % to split the pane side by side, and C-b " to split it top and bottom.
  3. Run claude, codex and gemini, one per pane.
  4. Detach with C-b d, and return with tmux attach -t agents.

The panes still share one checkout, and you still relay work between them by hand.

How do git worktrees keep the agents apart?

Give each agent that edits code its own worktree. A worktree is a second working directory for the same repository with its own branch checked out, so each agent edits its own files and commits to its own branch.

git worktree add -b agent/codex ../myapp-codex
git worktree add -b agent/gemini ../myapp-gemini

claude                            # plans and reads in the main checkout
cd ../myapp-codex && codex        # edits on agent/codex
cd ../myapp-gemini && gemini      # edits on agent/gemini

git worktree list shows every worktree and its branch. Git refuses to check out a branch that another worktree already has, which keeps two agents off the same branch. When a branch is merged, remove its worktree with git worktree remove ../myapp-codex.

Worktrees fix the files but not the hand-offs. The agents still cannot ask each other anything, and you still track which terminal is waiting for you.

How does MAO run all three together?

MAO puts each agent on a canvas as a terminal node that runs the agent's own CLI on its own login. You draw wires between the agents so they can send each other tasks, give each one a role, and put the agents that edit code on layers, which are git worktrees.

  1. Install and sign in to each CLI on the machine where its terminal will run. MAO runs the agent's own CLI and tells you when one is not installed. On Windows it finds agents installed with npm install -g, Gemini CLI among them, on your PATH.
  2. Open your repository as a workspace in MAO.
  3. Place the agents. Choose New terminal or agent, or press Ctrl+T (Cmd+T on macOS), then pick Claude Code and place its node. Repeat for Codex and Gemini CLI, which has its own tile in the agent sheet.
  4. Choose a login per node. Each node uses the CLI's own login unless you pick a saved account from its Account menu. Add account… saves a Directory, API key or Endpoint account, depending on what the provider supports.
  5. Give each a role. Builder and Reviewer are built in, for example Builder on Codex and Reviewer on Gemini CLI. Write a planning role for Claude Code in its instructions sheet and choose Save as role.
  6. Draw the wires. Select the Claude Code node, choose Connect and pick Codex, then connect Codex to Gemini CLI. A wire grants access only to the node at its other end. Wire a specification note to all three to share the plan.
  7. Give the builder a layer so its edits land on its own branch. You review the diff on the canvas and land the layer into Ground when it is ready.
  8. Ask once. Mark Claude Code as the Orchestrator in its setup sheet and tell it what you want built.

From its own terminal, the planner hands work to its peers with the mao command:

mao peers
mao ask builder "Implement the plan in the spec note"
mao ask reviewer "Review the builder's changes for bugs and missing tests"

mao ask sends a task and waits for the result. mao send returns a task id at once and mao wait <task-id> collects the result later. An Orchestrator can also add an agent with mao recruit <name> --preset <preset> --role <role> and wire it with mao connect <from> <to>.

MAO adds a role's instructions to prompts it sends over ACP. An agent that runs only as a terminal does not receive them automatically, so put the same rules in the project's instruction file too.

Which option should you use?

Start with the simplest one that covers your problem: terminals for a quick look, worktrees as soon as two agents edit code, and a canvas when you are tired of relaying work between them.

OptionSee all three at onceAvoids file collisionsAgents hand off workLogin per agentSetup
Separate terminalsNo, one window eachNoNo, you copy and pasteEach CLI's ownNone
tmux panesYes, panes in one windowNoNo, you copy and pasteEach CLI's ownInstall tmux
Git worktreesDepends on your terminalsYes, one branch eachNoEach CLI's ownOne git worktree add per agent
MAOYes, nodes on a canvasYes, with layersYes, wires and mao askThe CLI's own or a saved account per nodePlace nodes, draw wires

Frequently asked questions

Yes. Each is a separate CLI with its own login, so all three can run in one repository at once. Give every agent that edits code its own git worktree, or its own layer in MAO, so their changes do not collide.

No. Claude Code signs in with a Claude subscription and Codex with a ChatGPT plan, and Gemini CLI signs in with a Google account for organizations with a Code Assist licence. In MAO each agent runs on its own login, and MAO does not route your prompts through its own servers.

Gemini CLI 0.41 or later. With an older version, MAO says which version it found, where, and how to update it.

Yes, once you list it. Gemini CLI reads GEMINI.md by default; add AGENTS.md to the context.fileName setting and it reads that file too, the same one Codex and Claude Code read.

According to Gemini CLI's README, free individual access and Google AI Pro and Ultra moved to Antigravity CLI on June 18, 2026. MAO also runs Antigravity as an agent, beside Claude Code and Codex.

Sources

Checked on 2026-10-28.

Questions? Ask in the Discord or write to hello@miguer.com.

Try it on your own repo

Put your agents on one canvas.

A 7-day trial with every canvas feature. No card.

Download MAO