HhiveyBLOG

Codex and Claude Code Together: One Team, No Vendor Lock

You have Claude Code open in one tab and Codex in another. You copy a diff from one into the other and ask, "does this look right?" Then you paste the answer back. Most of your afternoon goes to carrying messages between two agents that can't talk to each other.

Running Codex and Claude Code together is worth it. Two different models catch different mistakes, and one can check the other's work instead of grading its own. But if you are the only link between them, you haven't built a team. You've become its switchboard.

This post covers a pattern you can use today with nothing but two terminals. After that I'll show how I run it in hivey, the tool I build.

Swarm commands made starting easy. Picking the agents is the new question

This week Google Antigravity announced a /swarm command. It orchestrates specialized agents in parallel, with a shared live message board and a choice of model per role. On the same day, Vorflux put the real problem plainly: "the bottleneck to parallel agents is initiating work, tracking it and getting approvals."

Taken together, those two posts mean starting a swarm is no longer the hard part. It's a slash command now. The harder questions are these:

A swarm built into one vendor's app runs on that vendor's terms. If your best reviewer comes from a different vendor than your best implementer, you want a coordination layer that doesn't belong to either one.

Why run Codex and Claude Code together

The strongest reason is independent review. An agent that checks its own work tends to agree with itself. It wrote the code from a set of assumptions, and it reads the code back with the same assumptions. A second agent, built on a different model with a different context, has no stake in the first agent's choices.

Other reasons:

The cost is coordination. Without structure, two agents produce two half-finished branches and a pile of contradictory notes.

A role split for Codex and Claude Code on one task

Give each agent a role before you give it a task. Here is the split I start from. Swap the agents freely; what matters is that the two roles stay separate.

RoleOwnsHands offNever does
Implementer (agent A)The branch and the code changesA diff plus a short note: what changed, why, what's untestedApprove its own work
Reviewer (agent B)Review notes and failing test casesA verdict: approve, or a numbered list of issuesEdit the implementer's files directly
Coordinator (you, or a lead agent)The goal and the definition of doneNext task, or "ship it"Write code during review

Three rules make this work without any special tooling:

  1. Hand off through files, not chat. The implementer writes HANDOFF.md and the reviewer writes REVIEW.md. You stop copying and pasting, and you get a record of every round.
  2. The reviewer can't fix things. If the reviewer edits code, you're back to self-review with extra steps. Issues go back to the implementer.
  3. Cap the rounds. After two or three review cycles without agreement, the coordinator decides. Agents arguing forever is still babysitting.

With two plain terminals, the loop looks like this: the implementer finishes and writes the handoff. You tell the reviewer to read the diff and the handoff and write a verdict. You pass the verdict back. It works, but notice who does every "you tell" and "you pass." You do.

Cross-review without becoming the messenger

To get out of the middle, the agents need three things you've been supplying by hand:

Add a coordinator that knows the goal and the role split, and you have a team you can step away from. That is the layer that shouldn't belong to a single model vendor.

How hivey runs Codex and Claude Code together

hivey is a terminal workspace and coordinating agent for your AI agents. You talk to hivey and it builds a swarm: a team with its own workspace and a dedicated coordinator that directs the agents, keeps the work moving and reports back to you. Teams can mix Claude Code and Codex in different roles.

Start where your code lives and describe the goal, including the role split:

hivey
hivey swarm new "Add rate limiting to the API. One agent implements, a second agent on a different model reviews every change before it's done."

Check what got built, which agents are on the team and what state they're in:

hivey swarm list
hivey swarm info <swarm>

The coordinator assigns the work and each agent runs in its own terminal space. Agents message teammates directly. Messages can go to an agent, a role or the whole swarm. Delivery is queued, and each agent has an inbox and a message history. That covers the handoff loop from the previous section without you in the middle.

To add context or change course, message the coordinator rather than an individual agent. Normal messages wait until the agent is idle:

hivey msg send <swarm>/coordinator "Reviewer: also check the rate limiter under concurrent requests."
hivey msg log --swarm <swarm>

The log lets you see who asked what, and when, without reading two transcripts.

You don't have to stay at your desk while this runs. Detach with ⌥Q (or ctrl+b, then q) and the agents keep working on your awake, connected machine. Run hivey to reattach. If you connect Slack, you can talk to hivey in its channel and to each swarm's coordinator in that swarm's channel.

For longer jobs, schedule a check-in so an idle coordinator picks the task back up:

hivey swarm schedule add <swarm> --every 2h "Check review status and report blockers."

If the job crosses a night, hivey quiet on silences sounds, voice and Slack posts while the agents keep going. When you turn it off, held Slack posts arrive as one catch-up message.

To see which setup providers are available on your machine, run:

hivey swarm providers

If you want to wire in your own provider, the repo includes a guide to writing one.

Try it

On an Apple Silicon Mac:

npm install -g @hivey/cli
hivey setup

On Linux or an Intel Mac, build from source from github.com/alyssaaa11/hivey. Then give hivey a goal that needs a builder and an independent reviewer, and let the coordinator carry the messages. Setup and everyday use are in the docs.