Clodex：把 Claude 塞进 Codex CLI 的 Rust 分支

Codex CLI with Claude stuffed inside. An abomination. Sam and Dario would not hold hands, so we did it for them. 🦞🤝🦞

Clodex is a fork of openai/codex where the model behind the ChatGPT desktop app's Codex is Claude, talking to Anthropic's Messages API in-process. No proxy, no translation layer, no TypeScript. The desktop app, the voice mode and your ChatGPT login stay exactly as they are; only the backend binary changes.

Built in one day, in one sitting, by voice. Treat it accordingly.

What it does

• WireApi::Anthropic: a new wire in codex-rs/anthropic that maps Codex's Responses-style prompt to Anthropic Messages (tools, thinking with signatures, prompt caching, images, native web search) and streams the result back as Codex events.
• Auth from Claude Code: reads ~/.claude/.credentials.json, refreshes the OAuth token in place, retries once on 401. Your Claude subscription is the quota.
• Voice mode keeps working: the realtime call is created on your ChatGPT account, the backend that answers is Claude. A voice-mode addendum is appended to every realtime session so Claude knows how to behave as a voice backend.
• Subagents wake the parent when they finish.
• Threads created under another provider resume on the active one.

Twelve commits on top of rust-v0.162.0-alpha.10. git log b496f984f4..HEAD is the whole story.

Heads up

• Tested on one machine: Arch Linux, ChatGPT desktop 26.930 (chatgpt-bin), Codex CLI 0.162.0-alpha.10 bundled. Nothing else.
• Needs an existing ChatGPT login in the desktop app (for the UI and voice) and an existing Claude Code login (for the model).
• Anthropic's OAuth is meant for Claude Code. Clodex presents itself as Claude Code. Whether that is inside the terms of your plan is your call.
• Unoptimized dev build, ~420 MB binary. No tests were run. No clippy. Expect rough edges: citations are dropped, apply_patch does not stream its diff, rate-limit headers are ignored.
• The desktop still runs the stock codex-code-mode-host; Clodex symlinks to it.

Install

git clone https://github.com/ZmoleCristian/clodex
cd clodex/codex-rs
RUSTFLAGS="-C link-arg=-fuse-ld=mold" cargo build -p codex-cli --bin codex --profile dev-small
ln -s /usr/lib/chatgpt/resources/codex-code-mode-host target/dev-small/codex-code-mode-host

Cold build is about 5 minutes on 8 cores; incremental builds are 30 to 50 seconds.

A model catalog with the Claude models is needed. Put it at ~/.codex/catalog.clodex.json (see docs/clodex-catalog.example.json).

Wrapper the desktop will launch instead of its bundled CLI:

#!/bin/sh
# ~/bin/clodex-codex
real=/path/to/clodex/codex-rs/target/dev-small/codex
case " $* " in
  *" app-server "*) exec "$real" "$@" \
    -c model_provider=anthropic \
    -c model=claude-opus-5-5 \
    -c model_catalog_json=$HOME/.codex/catalog.clodex.json ;;
  *) exec "$real" "$@" ;;
esac

Launcher:

#!/bin/sh
# ~/bin/clodex
pkill -f /usr/lib/chatgpt/ChatGPT 2>/dev/null
while pgrep -f /usr/lib/chatgpt/ChatGPT >/dev/null; do sleep 0.2; done
export CODEX_CLI_PATH=$HOME/bin/clodex-codex
exec /usr/lib/chatgpt/ChatGPT "$@"

The stock app, /usr/lib/chatgpt and ~/.codex are untouched. Start the app from its menu entry and you get stock Codex; start it with clodex and you get Claude.

Why

OpenCodex worked but it is a 480k-line TypeScript proxy between the app and the model. The seam in Codex is one match arm in core/src/client.rs; it was cheaper to put Anthropic there.

Upstream

This fork tracks openai/codex tags. Rebase, resolve the handful of conflicts in client.rs, realtime_conversation.rs and thread_processor.rs, rebuild. The upstream README is in docs/README.upstream.md.