Claude Code and Codex in one project

If you pay for both Claude and ChatGPT, you probably switch between Claude Code and Codex depending on the job, or the day, or which one still has room. We do. What we didn't like was running them as two separate worlds: two terminals, two sets of MCP servers, two sign-in dances.

So in Shuriken, one project can hold both. Each chat picks its agent, and everything around the chat is shared.

Each chat picks its agent#

When you start a chat, the model picker lists the agents your computer can run first. Pick Claude Code for one chat and Codex for the next. They sit side by side in the same project, in the same sidebar.

Every chat (apart from the Main chat, which works on your own branch) gets its own branch in its own git worktree. So a Claude Code chat and a Codex chat can work at the same time without stepping on each other's files. When one is done, Finish runs your checks and merges it, or opens a pull request. The parallel agents page explains the worktree setup.

One set of MCP servers for both#

This is the part that used to hurt. Claude Code keeps its MCP servers in its own JSON files (~/.claude.json or a project's .mcp.json). Codex keeps them in ~/.codex/config.toml, one [mcp_servers.<name>] table each. Two formats, two places, and two sign-ins to Linear or Notion.

In Shuriken you add a server once, in Global MCPs, and assign it to every project or just some. Claude Code and Codex chats are both handed the same servers when they start. The sign-in belongs to the server, not to your AI account, so it survives switching accounts, projects and agents. There are 57 ready-made connectors, and any other server can be added by address or command.

Each agent keeps its own sources#

Claude Code chats run on your Claude sources: subscription sign-ins and Anthropic keys. Codex chats run on your ChatGPT sign-ins and OpenAI keys. They don't borrow from each other.

Switching works per agent, too. When the Claude account in use hits your switch point, Claude Code chats move to another of your Claude accounts on their next request. When a ChatGPT account runs out, Codex chats move to another of yours when the current turn ends. You don't log in or out of anything.

Even the Free plan covers the basic setup: two sources, so one Claude and one ChatGPT.

Ideas for splitting the work#

We're not going to tell you which agent is better. That changes with every model release, and it depends heavily on your code. (Our Claude Code vs Codex comparison lays out the differences without picking a winner.) Here are some ways to split work that we've found useful:

• Same task, two chats. Give the same prompt to a Claude Code chat and a Codex chat. Each works on its own branch. Compare the two diffs in Changes and finish the one you like. It costs double, so save it for the decisions that matter.
• One writes, the other reviews. Worktrees share the repository's branches, so a Codex chat can read the branch a Claude Code chat made. Ask it for a review before you finish.
• Split by room. If one provider's week is getting tight, send the routine jobs (tests, docs, small fixes) to the agent with more room, and keep the other for the hard parts.
• Split by taste. Some people prefer one agent for planning and the other for grinding through a long list of small changes. Try both for a week and see where you end up.
• Keep the queue moving. Type the next task into either chat while it works. Queued messages wait their turn.

Not just these two#

Claude Code and Codex are the two most people ask about, but Shuriken runs 67 coding agents, including Gemini CLI, GitHub Copilot, OpenCode and Cursor. The same rules apply to all of them: one project, a chat per agent, shared MCP servers. The most detail (Activity, switching mid-task) comes from Claude Code.

Want both agents in one place? Try Shuriken and add a source for each.