claudex

Run Claude Code over your ChatGPT subscription.

claudex is a local proxy that translates Anthropic Messages API to the OpenAI Responses API used by the Codex backend, and authenticates with your ChatGPT Plus/Pro account via the same OAuth device flow the codex CLI uses. Your Claude Code stays unchanged — it just points at ANTHROPIC_BASE_URL=http://127.0.0.1:<port> and talks to the proxy.

Status: works for me, not battle-tested. SSE event shape from the Codex backend drifts; expect rough edges around tool-use and reasoning streams.

Install

cargo install --git https://github.com/m0n0x41d/claudex

or from a clone:

cargo install --path .

Use

claudex auth                      # one-time: ChatGPT device-flow login
claudex init                      # write default config to ~/.claudex/config.yaml
claudex                           # spawn the proxy and run `claude` through it
claudex status                    # show current token state
claudex --resume <session_id>     # any unknown args are forwarded to `claude`
claudex -- --print "hello"        # use `--` to be explicit

Arg order to claude: claude.default_args from config, then whatever you passed to claudex.

claudex with no subcommand binds an ephemeral port, exports ANTHROPIC_BASE_URL + a dummy ANTHROPIC_API_KEY, then execs claude with whatever args you pass after it. When claude exits, the proxy is dropped. No daemon, no persistent ports.

Logs go to ~/.claudex/claudex.log while claude is running (stderr is left alone so the TUI stays clean). tail -f ~/.claudex/claudex.log to watch traffic.

If claude is already signed in to claude.ai, run claude /logout inside it once and answer No to the API key approval prompt — otherwise the claude.ai token takes precedence over the proxy.

Config

~/.claudex/config.yaml:

proxy:
  bind: 127.0.0.1
  port: 0                # 0 = ephemeral
  log_level: info
codex:
  base_url: https://chatgpt.com/backend-api/codex
  default_model: gpt-5.5
  reasoning_effort: medium      # low | medium | high
  reasoning_summary: concise    # auto | concise | detailed
  originator: claudex
claude:
  command: claude
  default_args: []
model_map: {}                   # claude-opus-4-7: gpt-5.5

Available models depend on your ChatGPT tier — the Codex backend gates them by client_version. Probe with:

curl -s -H "Authorization: Bearer $(jq -r .access_token ~/.claudex/auth.json)" \
     -H "chatgpt-account-id: $(jq -r .account_id ~/.claudex/auth.json)" \
     'https://chatgpt.com/backend-api/codex/models?client_version=1.0.0' | jq '.models[].slug'

Architecture

Layered, bottom-up. Each layer depends only on the layer below.

types/      L0  algebraic data (Anthropic + Responses wire types, TokenSet, Config)
translate/  L1  pure Anthropic↔Responses (incl. StreamReconciler state machine)
auth/       L2  pure auth state machine (classify, device-flow parsing)
ports/      L3  effect traits + their concrete impls (http, fs, subprocess, …)
app/        L4  use cases: authenticate, AuthService, serve, launch_session
cli/        L5  argv dispatch

Side effects live only in ports/ and app/. Translation and auth state are pure and unit-tested. The proxy is spawn-per-invocation: one claudex process = one claude session = one ephemeral listener.

License

MIT.