Skip to main content

Embedding Claude Code and Codex in the same workspace

Status: Work in progress. Notes from my own experiments -- expect revisions.

Both Claude Code (Anthropic) and OpenAI Codex CLI are useful AI coding tools, and they're not mutually exclusive. This is a working doc on running them side by side in the same project directory.

Why both​

  • Different model strengths. Sometimes one nails a problem the other muddles.
  • Different ecosystems. Codex integrates with OpenAI Agents SDK; Claude Code has the Agent SDK and a richer skills/memory system.
  • Diversification. If one platform has an outage or limit hit, the other keeps you working.

Installation​

Both are installed independently:

# Claude Code
npm install -g @anthropic-ai/claude-code

# OpenAI Codex CLI
npm install -g @openai/codex

Each has its own auth flow and config directory:

ToolConfig dir
Claude Code~/.claude/
Codex~/.codex/

No conflicts at the system level.

Pointing both at the same project​

Just open two terminals in the project folder:

# Terminal 1
cd ~/code/myproject && claude

# Terminal 2
cd ~/code/myproject && codex

Both read and write the same files. The conflict surface is the filesystem, not the tools themselves.

Avoiding stepping on each other​

The real risk is both tools editing the same file in the same session and one overwriting the other's work. A few patterns:

  • Single-active-tool rule. Only one tool actively makes edits at a time. Use the other for read-only Q&A.
  • Git as the reconciliation layer. Commit frequently. If both tools touch overlapping code, you see the conflict on the next commit.
  • Different branches per tool. Spin up worktrees: claude on feature/x-claude, codex on feature/x-codex. Compare and merge manually.

Configuration files: separate or shared?​

CLAUDE.md is read by Claude Code. Codex looks for AGENTS.md or .codex/instructions.md.

You can either:

  1. Keep them separate. Different files, different audiences (each tool gets its own personality).
  2. Symlink one to the other (or a shared AGENTS.md that both reference). Single source of truth, but you lose tool-specific instructions.

I'm currently running them separate, with about 80% content overlap that I copy-paste when something general changes. Not ideal -- looking for a better pattern.

Open questions (things I haven't figured out yet)​

  • Best workflow for handing off context between the two tools mid-task.
  • Whether to share or isolate session memory (Claude Code has explicit memory; Codex has more session-local state).
  • How to make MCP servers configured for one available to the other.

If you've solved any of these, open an issue or PR -- I'll update this page.