Use Claude Code as a model in the Codex desktop app.

Tandem adds Claude Opus 5.5 (Claude Code) to the Codex model picker, next to your GPT models. Your own Claude Code installation runs the agent loop (its prompt, read and search tools, web tools, skills, subagents, MCP servers and CLAUDE.md), while Codex runs the shell commands and file edits with its normal approvals and shows them as its own command and diff cards. GPT keeps working exactly as before.

Warning

Tandem is an independent, experimental project, not affiliated with or endorsed by Anthropic or OpenAI. It uses your Claude Code installation and your Claude login or API key, so your agreement with Anthropic governs that usage. Read Terms and account risk before using it with a Claude subscription.

Tip

Why Claude Code runs as the harness, instead of plugging Opus into Codex directly

Tandem's first version treated Opus as a bare model behind Codex. It was dropped because:

• Your Claude login sat in the middle. A local relay forwarded your login token and edited Claude Code's requests. Now the unmodified CLI talks straight to Anthropic and Tandem never sees your token.
• It used Claude Code as a model API. Tandem runs Claude Code as Claude Code, which is much closer to ordinary use. It is still a third-party front end, so the warning above applies.
• It cost far more. Codex sends every tool definition it has with each request:

Either way, Codex runs every command and edit, and GPT requests and OpenAI credentials never reach Claude Code.

How it works

Codex desktop ──► local router ──┬─ GPT ─────────► OpenAI (unchanged request, your OpenAI login)
                                 └─ Claude Code ─► local bridge ─► claude -p (your Claude login)
                                                                     └─ commands and edits ─► back to Codex as native tool calls

Enabling Tandem points Codex's openai_base_url at a local router and adds one model to the catalog Codex reads. The router passes GPT traffic straight to OpenAI. The Claude Code model goes to a local bridge that keeps one Claude Code session per chat. Claude Code's own shell and edit tools are turned off; it gets Codex's exec_command, write_stdin and apply_patch instead, and each call is handed to Codex to run, then its result goes back to Claude Code.

Requirements

• Windows 10/11 (the only tested platform)
• Codex desktop, signed in with ChatGPT, plus the Codex CLI for installing (tested: app 26.924, CLI 0.157)
• Claude Code, signed in with access to Claude Opus 5.5 (tested: 2.1.283)
• Python 3.11+ with cryptography and Node.js 22.15+. Codex desktop's bundled runtimes are used automatically.

Install

codex plugin marketplace add Yuruzuu/tandem
codex plugin add tandem@tandem

Then open Codex, ask "Enable Tandem" and approve the tool call. Fully restart Codex and pick Claude Opus 5.5 (Claude Code). Your GPT default stays selected for new chats.

Enabling changes only openai_base_url and model_catalog_json in config.toml and saves a backup. It refuses to run if you already use a custom provider or base URL.

Good to know

• Approvals: commands and edits follow your Codex approval and sandbox settings. Claude Code's other tools (reads, web, its own MCP servers) run without Codex approval.
• Not shared: Codex's developer instructions, skills, plugins and MCP servers. Claude Code uses its own setup; your AGENTS.md and environment context are passed along.
• Switching models mid-chat works both ways, except after compaction: Claude Code can't read GPT's compacted history, and GPT only sees a pointer to a compacted Claude Code session. Start a fresh chat for those.
• Sessions resume automatically after interruptions and restarts, and idle ones stop after an hour.
• Activity panel: commands, edits and web searches appear as Codex's own rows. Files Claude Code reads don't, since Codex only lists reads it ran itself, and like GPT's, Claude's thinking shows live and is hidden once the turn ends.
• Effort in Codex maps to Claude Code's --effort (low to max).

Uninstall and recovery

While Tandem is enabled, all Codex model traffic goes through its router, which the plugin starts when Codex loads. If Codex can't reach any model, or to uninstall, restore the direct route from a clone of this repo (works without Codex or the router running):

git clone https://github.com/Yuruzuu/tandem
.\tandem\plugins\tandem\scripts\tandem.cmd restore-desktop   # or: start / status / stop
codex plugin remove tandem@tandem
codex plugin marketplace remove tandem

Restart Codex afterwards. Keep %USERPROFILE%\.codex\tandem if you might reinstall: it holds the key for Claude Code turns in existing chats, plus your config backups.

Environment variable	Purpose
TANDEM_HOME	Private state folder (default %USERPROFILE%\.codex\tandem)
TANDEM_PYTHON, TANDEM_NODE, TANDEM_CLAUDE	Override the Python, Node.js or claude executable

Security

Everything runs on loopback (127.0.0.1:57855 and :57856) behind random secrets, and browser requests are rejected. GPT requests go unchanged to the fixed chatgpt.com/backend-api/codex endpoint; the only edit is removing Tandem's own encrypted items when a chat switches from Claude Code to GPT. Tandem never reads your OpenAI or Claude credential files. Report vulnerabilities privately: see SECURITY.md.

Terms and account risk

Tandem runs the official, unmodified Claude Code with your own login or API key, billed to your own account, and never handles your Claude credentials. Anthropic's legal and compliance page describes subscription (Free, Pro, Max) sign-in as intended for ordinary use of Claude Code and Anthropic's own apps, and says it may enforce its restrictions without notice. Using a subscription through another app's interface may fall outside that, and you are responsible for deciding whether your use complies. For certainty, sign Claude Code in with an Anthropic API key. OpenAI's terms continue to apply to the GPT traffic Tandem forwards.

Development

cd plugins\tandem
python -B -m unittest discover -s tests -p "test_*.py" -v
$env:PYTHON = "python"; node --test tests/router.test.cjs

The tests run offline with a fake claude process. Main pieces: scripts/router.cjs (router), scripts/bridge.py and scripts/harness.py (Claude Code sessions and the Codex relay), scripts/control.py (setup and restore). See CONTRIBUTING.md.

Acknowledgements and license

Helpers in scripts/native/cli.py are adapted from Nous Research's hermes-plugin-claude-subscription-directsdk (MIT, see its license). Tandem is released under the MIT License.