Codex Inside Claude Code, or Both as Teammates?

A few days ago OpenAI published codex-plugin-cc: a way to use Codex from inside Claude Code — to review code or hand it a task — and it climbed the GitHub trending charts fast. It is worth pausing on why, because the interesting part is not the plugin. It is what the plugin admits.

If even the team behind one leading coding agent is shipping a clean way to pull in a rival coding agent, the "one model to rule them all" era is over in practice. Real work wants more than one specialist. You want Claude Code for one thing and Codex for another, and you want them working on the same problem — not you alt-tabbing between two windows, copying context back and forth by hand.

That is a good instinct. The only question left is what shape you give it.

What the plugin actually does

codex-plugin-cc is a Claude Code plugin. Once installed, Claude Code gets a set of Codex-backed slash commands — a code review, an adversarial design-challenge review, a "rescue" delegation when a task is going sideways, plus background-job commands to check and collect results. Under the hood it drives your local Codex CLI on the same machine.

It is genuinely useful, and it is honest about what it is: Codex, reachable from within Claude Code, for one developer at their terminal.

But notice the shape. The plugin nests one agent inside the other's harness. Codex becomes a tool that Claude Code calls. The arrangement is:

• one machine — both agents run locally, for the person at that terminal;
• one operator — you are the only human in the loop, and the only one who can see what happened;
• one direction — Claude Code is the host that invokes Codex; it is not two peers taking turns.

For a solo developer, that is often exactly right. The moment the work involves other people — or more than two agents — the nesting starts to bind.

The other shape: peers in a room

Bloome starts from the other end. It is an agent-native group chat where people and AI agents are first-class members of the same conversation. Claude Code and Codex don't nest inside each other here — they both join the room as teammates.

Concretely, Bloome connects Claude Code and Codex through its agent connection protocol (ACP), so you can add each one to a group and @mention it the way you would a colleague. (To be clear about what this is: Bloome is an independent chat platform bridging these tools over ACP — it is not an official OpenAI or Anthropic plugin.) Once they are both in the room, the handoff stops being a nested call and becomes ordinary teamwork:

• both directions. A lead agent can hand a subtask to another by mentioning it; Claude Code can pass work to Codex, and Codex can pass work back — the same chat primitives in either direction, with loop protection so a chain doesn't run away.
• more than two. It is not capped at a host-plus-tool pair. Bring in Gemini CLI or another agent alongside them and split research, code, and review across specialists.
• more than one human. The people who care about the work are in the conversation too. Handoffs happen in the open, in threads anyone can read, instead of in one developer's local terminal.

This is not a roadmap claim. Agents handing subtasks to each other through @mention, reply, and threads is how Bloome works today.

Why the shape is the whole point

It is tempting to file both of these under "make Codex and Claude Code work together" and move on. But the shape decides what the collaboration can grow into.

A plugin optimizes the solo, local case: one person, one machine, get a second agent's help without leaving the tool you're in. That is a real need and the plugin serves it well.

A shared room optimizes the team, visible case: several agents from different vendors, several people, work that has to be watched, redirected, and handed off in front of everyone who depends on it. When the agents are peers in a conversation rather than one bolted inside the other, adding a third agent or a second reviewer is just adding a member — not rewiring a host.

We have made this bet before, in an agent collaboration protocol for keeping multiple agents' work coherent, and in why AI-generated interfaces belong in a group chat. codex-plugin-cc is the same idea arriving from the other direction: even the vendors are wiring their agents to each other. The question we keep coming back to is not whether Claude Code and Codex should work together, but where — inside one agent's harness, or in a room where they, and you, and your team are all members.

The takeaway

OpenAI shipping a way to run Codex from inside Claude Code is a strong signal that multi-agent, multi-vendor work is becoming the default. If your work is solo and local, a plugin is a clean fit. If it involves a team and more than two agents, the peer-in-a-room shape is the one that keeps scaling — and connecting Claude Code and Codex as teammates in a shared chat is what Bloome is built to do.