Guide

How to use Claude Code and Codex together

In short

You can run Claude Code and Codex on the same project. The simplest setup gives both the same written context, lets 1 agent write and the other review, and checks the result before you read it. brainplane does this on your Mac with your own plans.

Updated 9 October 2026. Checked with brainplane 0.2.2.

Download for Mac

14 days free with every feature. Email only, no card.

Why use both

Claude Code and Codex are 2 command line coding agents from 2 different model families: Anthropic for Claude Code, OpenAI for Codex. Many developers pay for both plans and use them on the same repo. A common split is 1 agent that writes and 1 that reviews. The idea: a model from another family reads the work with fresh eyes.

The hard part is the context. Each agent starts a session knowing nothing about the decisions you made in the other one. So you copy the goal, the constraints and the last answers from 1 terminal to the other.

3 ways to combine them

Way	How it works	Good for	What it costs you
Copy and paste	You brief each agent by hand and paste results across	1 short task	Your time on every round. Context drifts
Shared files in the repo	Both read the same notes, such as an `AGENTS.md` and a plan file	Stable project rules	You keep the files current. Answers and decisions get lost between sessions
A shared brain with handoffs	A local service keeps the topic's state, briefs each agent and runs checks on the result	Daily work with 2 or 3 agents	1 app on your Mac

brainplane is the third way. It gives both agents 1 shared brain per topic and moves work between them under a contract.

Set it up with brainplane

Install brainplane. Download it, drag it to Applications and start the 14-day trial with your email.
Connect your agents. Setup finds Claude Code and Codex on your Mac and adds 1 MCP server entry named brainplane to each, with a backup of each file it changes.
Work in Claude Code as usual. Through the MCP server it reads the topic's brief and logs its decisions to the topic.
Hand the result to Codex for review. Codex runs read-only in its own sandbox, with a context pack of 3,000 tokens at most.
Read 1 finished card. The card shows the verdict, the checks, where the result is and the command to resume the session.

The handoff from step 4, from the command line:

bp delegate codex --mode review "Review the payment retry change in src/billing for race conditions"

For a quick second opinion on 1 to 6 files, the reviewer defaults to the other family:

bp second-opinion src/billing/retry.ts --agent codex

What each agent receives

Each agent gets a context pack for the topic: the goal, where things stand, settled decisions, answered questions, people and files. Binding items come first, older notes are trimmed first. After the first pack, an agent gets only what changed since its last one. You can see the exact pack before it goes:

bp pack <topic> --for codex

What stays the same

brainplane runs the official, unchanged tools: claude -p and codex exec. They keep their own logins, updates and settings. Usage counts against your own plans, like any session you start yourself. brainplane never reads their login files or tokens.

Questions

Can Claude Code and Codex work on the same repo at the same time?

Yes. brainplane records which session touched which file, lets a session claim a file for 30 minutes and flags it when another live session writes the same file in that time. It does not lock files on disk, so many people also give each agent its own git worktree.

Do I need an API key for Codex or Claude Code?

No. Both run on the login you already use with them. brainplane never asks for an agent API key.

Does brainplane call Claude or OpenAI models itself?

The background service makes no model call of its own. Agents run when you start a run or press Run on a card. 2 options you turn on are separate: OpenAI speech on your own key, and 1 short sorting call on your own Claude Code or Codex when brainplane cannot place a request locally. You can turn sorting off.