claude-code-chatgpt-proxy

Run Claude Code on your existing ChatGPT Plus/Pro/Max subscription. No API key, no separate billing. Your normal claude command stays untouched: you source one env file in a shell, and claude in that shell routes through your ChatGPT quota.

The mechanism is LiteLLM running as a local proxy. Claude Code speaks the Anthropic API to 127.0.0.1:4001; LiteLLM forwards to OpenAI's ChatGPT subscription backend using OAuth device-flow auth.

Based on the setup described in Run Claude Code with Your ChatGPT Subscription by Prabal Gupta, repackaged for Linux with uv instead of conda, with the LiteLLM compatibility patch applied automatically and the port moved off 4000 (a common conflict). See docs/how-it-works.md for what each piece does.

What you get

you type:  claude "fix this test"
           |
           v  Anthropic API format
LiteLLM proxy (127.0.0.1:4001)
           |
           v  ChatGPT subscription backend (your OAuth login)
gpt-5.5 / gpt-5.3-codex / gpt-5.6-sol

• No credentials in this repo, ever. The auth key is generated on your machine at setup time; ChatGPT OAuth tokens are stored by LiteLLM locally.
• Pinned litellm==1.83.0. Versions 1.82.7 and 1.82.8 were compromised with credential-stealing malware (2026-03-24); 1.83.0 is the first release from the hardened CI/CD v2 pipeline. setup.sh verifies the installed version.

Requirements

• Linux or macOS
• uv on PATH
• openssl, patch (standard on most systems)
• Claude Code installed
• A ChatGPT Plus, Pro, or Max subscription

Quick start

Step 1. Clone and run setup:

git clone https://github.com/9prodhi/claude-code-chatgpt-proxy.git
cd claude-code-chatgpt-proxy
./setup.sh

You should see patch OK, litellm 1.83.0, and Setup complete.

Step 2. Start the proxy and complete the one-time ChatGPT login:

./scripts/litellm-proxy.sh

The terminal prints a device code:

Sign in with ChatGPT using device code:
1) Visit https://auth.openai.com/codex/device
2) Enter code: XXXX-XXXX

Open the URL, log in with your OpenAI account, enter the code. The proxy's port does not open until this login finishes ("Waiting for application startup" is expected, not a hang). After login you'll see Uvicorn running on http://0.0.0.0:4001. Tokens are saved locally; future starts don't prompt.

Step 3. In another terminal, run Claude Code through it:

source scripts/claude-codex-env.sh
claude

That's it. claude in this shell bills your ChatGPT subscription. claude in any other shell is your normal Anthropic account.

Model mapping

What Claude Code's model tiers resolve to (set in scripts/claude-codex-env.sh):

Claude Code tier	ChatGPT model	Notes
default / Sonnet	gpt-5.5	Strong all-rounder, works on Plus
Haiku (fast)	gpt-5.3-codex-spark	Fast code tasks, works on Plus
Opus / Fable	gpt-5.6-sol	Deep reasoning. May require Pro/Max. 30+ min thinking is normal

Switch per-session: claude --model gpt-5.3-codex.

Model names drift. If your ChatGPT model picker shows different names, edit both model_name and model in ~/.litellm/chatgpt-config.yaml and restart the proxy.

Configuration

All knobs are environment variables read by setup.sh and both scripts:

Variable	Default	Effect
VENV_DIR	<repo>/.venv	Where the Python venv lives
LITELLM_DIR	~/.litellm	Config, master key, OAuth token storage
PROXY_PORT	4001	Proxy listen port. Keep off 4000 (NoMachine and other tools squat it)

Files on your machine after setup:

Path	Purpose	In repo?
~/.litellm/chatgpt-config.yaml	Model routing config	template in config/
~/.litellm/master-key.txt	Local auth key between Claude Code and the proxy	never
<repo>/.venv/	LiteLLM install (patched)	gitignored

Known limitations

Claude Code's harness (system prompts, tool-use patterns, agent loops) is built for Claude models. Most GPT models handle general coding fine through this setup, but agentic multi-step workflows may behave differently than with Claude. The Spark model in particular does not play well with the harness. It works; don't expect parity.

Rate limits come from your subscription tier: Plus has lower ceilings, Pro and Max suit heavy agentic use.

Troubleshooting

See docs/troubleshooting.md. The two you'll most likely hit:

• Unable to connect to API (ECONNRESET): something else owns your port, or the proxy is still waiting for OAuth. Details in the doc.
• can only concatenate list (not str) to list: the LiteLLM patch didn't apply. Re-run ./setup.sh and check for patch OK.

Security notes

• Nothing in this repo contains or transmits credentials. Verify yourself: the master key is generated by openssl rand in setup.sh, OAuth tokens are handled by LiteLLM's ChatGPT provider on your machine.
• The master key only authenticates localhost traffic between Claude Code and your own proxy. Treat ~/.litellm/ as private anyway (master-key.txt is 600).
• After any LiteLLM upgrade, re-check the supply-chain advisory before unpinning 1.83.0, and re-apply the patch if the upstream fix hasn't landed (./setup.sh is idempotent and safe to re-run).

License

MIT. See LICENSE.