claudecodex lets you use Claude Code with OpenAI models — GPT-5.6 (Sol / Terra / Luna) as the engine — billed to your ChatGPT Plus/Pro subscription through a local Codex proxy, without an Anthropic API key or an OpenAI API key.
You keep the Claude Code agent workflow you like (file edits, terminal flow, approvals, MCP servers, skills) but the model answering is GPT-5.6, running on the ChatGPT subscription you already pay for instead of your Anthropic usage.
One-line install. Plain claude stays completely stock.
claudecodex launches the real Claude Code UI, points it at a local Anthropic-compatible proxy on 127.0.0.1:18765, and the proxy forwards requests to GPT-5.6 via Codex auth on your ChatGPT subscription:
claudecodex
→ claude (real Claude Code UI)
→ local Anthropic-compatible proxy (127.0.0.1:18765)
→ Codex auth via your ChatGPT Plus/Pro subscription
→ GPT-5.6-sol
→ streamed back into Claude Code
Jump to: Install · Usage & model switching · Configuration · vs alternatives · Troubleshooting · FAQ: Claude Code without an Anthropic API key
(8× speed — watch the full demo video)
- Claude Code UI + GPT models — the best agentic coding interface, driving
OpenAI's strongest coding model (GPT-5.6 Sol at
maxreasoning effort). - Uses the ChatGPT subscription you already have — no pay-per-token API bills; requests go through Codex auth on your Plus/Pro plan.
- Zero interference — a thin launcher that
execsclaudewith process-local environment variables. Your normalclaudecommand, login, settings, and Anthropic models are untouched. - Full 372K context window — configured to GPT-5.6 Sol's real Codex context limit, so the context gauge and auto-compaction are accurate.
- Model switching built in — flip between Sol (best), Terra (mid), and Luna (fast/cheap) per launch or mid-session.
claudecodex is a thin launcher: it starts the proxy if needed, exports the
right environment only for that process, then execs claude.
- Claude Code — the
claudeCLI. Install: https://claude.com/claude-code - claude-code-proxy — the local Anthropic⇄Codex proxy by @raine. Needs a build that knows the GPT-5.6 models (v0.1.8+; v0.1.17 or newer recommended).
- A ChatGPT Plus/Pro account, authenticated through the proxy's Codex login.
The proxy keeps its own login (it does not reuse the standalone Codex
CLI's credentials), so you authenticate it once with
claude-code-proxy codex auth login.
This is for personal, local use with your own subscription. Don't expose it as a public API or resell access — routing a ChatGPT subscription through non-official clients is a gray area under OpenAI's terms.
1 — install the launcher (one-liner):
curl -fsSL https://raw.githubusercontent.com/karem505/claudecodex/main/install.sh | bash2 — install the proxy (skip if you already have v0.1.8+):
curl -fsSL https://raw.githubusercontent.com/raine/claude-code-proxy/main/scripts/install.sh | bash3 — log the proxy in to your ChatGPT/Codex account:
claude-code-proxy codex auth login # browser OAuth
# or, on a headless box:
claude-code-proxy codex auth device # device-code flowThat's it — the installer checks steps 2 and 3 for you and prints these exact commands if anything's missing.
claudecodex # start Claude Code on gpt-5.6-sol (max effort)
claudecodex --mini # start on gpt-5.6-luna (fast/cheap) (alias: --luna)
claudecodex --terra # start on gpt-5.6-terra (mid tier)
claudecodex --big # start on gpt-5.6-sol explicitly (alias: --sol)
claudecodex "fix the bug" # any normal claude args pass straight throughSwitch models mid-session: type /model, pick gpt-5.6-sol or gpt-5.6-luna,
then press s (session-only). Both models are always listed. Use s, not
Enter — Enter saves it as a default in the shared settings and would change
plain claude too.
By default claudecodex starts in --dangerously-skip-permissions (bypass)
mode so tool calls run without prompting. Toggle it off with the env var below,
or cycle modes live with shift+tab.
gpt-5.6-sol launches at max effort — the highest reasoning level reachable
through claude-code-proxy today (forwarded to Codex as reasoning.effort=max).
luna/terra keep Claude Code's own effort setting. Change it live with
/effort, or per-launch with CLAUDECODEX_EFFORT.
About "ultra": GPT-5.6 Sol's
ultramode (parallel subagents) exists in the official Codex app, but it is not yet exposed by claude-code-proxy — the proxy caps atmax, and Claude Code's effort scale has noultra(it would clamp toxhigh).claudecodextherefore usesmax, the real ceiling. If you setCLAUDECODEX_EFFORT=ultrait is auto-downgraded tomaxwith a notice; once a proxy build shipsultra, that env value will pass straight through.
All optional — set as environment variables:
| Variable | Default | Meaning |
|---|---|---|
CLAUDECODEX_MODEL |
gpt-5.6-sol |
The "big" model |
CLAUDECODEX_SMALL_MODEL |
gpt-5.6-luna |
The fast/cheap model |
CLAUDECODEX_EFFORT |
max (on sol) |
Thinking level: low/medium/high/xhigh/max |
CLAUDECODEX_CONTEXT |
372000 |
Context window (see note) |
CLAUDECODEX_SKIP_PERMISSIONS |
1 |
1 = bypass permissions, 0 = normal prompts |
CLAUDECODEX_PORT |
18765 |
Proxy port |
CLAUDECODEX_AGENT_VIEW |
0 |
1 = show the background-jobs dashboard on start |
CLAUDECODEX_PROXY_BIN |
claude-code-proxy |
Proxy binary name/path |
Example:
CLAUDECODEX_EFFORT=high CLAUDECODEX_MODEL=gpt-5.6-terra claudecodexgpt-5.6-sol's Codex context window is 372K tokens (~353K effective after
the 95% multiplier). The API model is 1.05M, but the Codex subscription path caps
at 372K — advertising more triggers "exceeds the context window" errors. The
272K figure often quoted is just the pricing knee (input past 272K bills at
~2×), not the hard cap. Claude Code otherwise assumes 200K for unknown model
names, so claudecodex sets the real 372K; its autocompact buffer keeps a margin
below that. Tune with CLAUDECODEX_CONTEXT.
Everything below is exported inside the launcher process and dies with it —
nothing is written to ~/.claude/settings.json or your shell:
ANTHROPIC_BASE_URL=http://localhost:18765
ANTHROPIC_AUTH_TOKEN=claudecodex-local
ANTHROPIC_MODEL=gpt-5.6-sol
ANTHROPIC_SMALL_FAST_MODEL=gpt-5.6-luna
CLAUDE_CODE_MAX_CONTEXT_TOKENS=372000
CLAUDE_CODE_AUTO_COMPACT_WINDOW=372000
CLAUDE_CODE_DISABLE_AGENT_VIEW=1
# plus, session-scoped via flags: --model / --effort / --dangerously-skip-permissions
Your plugins, skills, and MCP servers are shared (the launcher does not
change CLAUDE_CONFIG_DIR), so they work in claudecodex too. Note that
claude.ai-hosted connectors are disabled while a token auth source is set —
that's expected for any custom endpoint.
Install the claudecodex launcher, install claude-code-proxy
(v0.1.8+), and run claude-code-proxy codex auth login once. Then run claudecodex —
Claude Code starts with GPT-5.6 Sol as the model, billed to your ChatGPT subscription.
See Install.
Yes — that is exactly what claudecodex does. Claude Code talks to any
Anthropic-compatible endpoint via ANTHROPIC_BASE_URL;
claude-code-proxy translates
those requests to OpenAI's Codex backend, so Claude Code runs GPT-5.6
(Sol, Terra, or Luna) as its model.
Yes. In a claudecodex session, no Anthropic account is used at all — auth and
billing go through your ChatGPT Plus/Pro subscription via Codex. (You still
install the Claude Code CLI itself, which is free to download.)
No. Requests are billed against the ChatGPT Plus/Pro plan you already have — there are no per-token API charges, because it uses Codex subscription auth, not an OpenAI API key.
No. The launcher exports everything process-locally and passes session-scoped
flags only. Plain claude keeps its own login, model, settings, and history.
See What it sets.
- Sol (
claudecodex) — strongest coding model, launches atmaxeffort. Default. - Terra (
claudecodex --terra) — mid tier, faster. - Luna (
claudecodex --mini) — fastest/cheapest, good for quick edits.
It's for personal, local use with your own subscription. Routing a ChatGPT subscription through non-official clients is a gray area under OpenAI's terms — don't expose it as a public API or resell access. Use at your own discretion.
| claudecodex | claude-code-router / LiteLLM | OpenRouter bridges | |
|---|---|---|---|
| Billing | ChatGPT Plus/Pro subscription — no per-token cost | OpenAI API key, pay per token | OpenRouter credits, pay per token |
| Anthropic API key needed | No | No | No |
| Setup | One-liner + one proxy login | Router/proxy config files | API key + config |
Touches plain claude |
No — process-local env only | Often global config/env | Varies |
GPT-5.6 Sol at max effort |
Yes, default | Depends on provider support | Depends on provider support |
If you want pay-per-token API access with many providers, a router is the better fit. If you want the strongest OpenAI coding model inside Claude Code on the flat-rate ChatGPT subscription you already pay for, that's claudecodex.
- raine/claude-code-proxy — the local Anthropic⇄Codex proxy that makes this possible.
- Claude Code — Anthropic's agentic coding CLI.
- OpenAI Codex — the subscription backend serving GPT-5.6.
The Codex subscription path caps GPT-5.6 Sol at 372K tokens (the API model's
1.05M does not apply). claudecodex sets 372K by default; if you overrode
CLAUDECODEX_CONTEXT, lower it back to 372000.
Check that claude-code-proxy is installed (v0.1.8+) and authenticated:
claude-code-proxy --version
claude-code-proxy codex auth loginIf the proxy fails to start, check the log at
~/.local/state/claudecodex/proxy.log.
Proxy upgrades can reset its Codex login (it's separate from the standalone
Codex CLI's login). Re-run claude-code-proxy codex auth login.
rm ~/.local/bin/claudecodex
# optional: remove the proxy + its login
claude-code-proxy codex auth logoutMIT — see LICENSE.
Not affiliated with Anthropic or OpenAI. "Claude Code" and "GPT" are trademarks of their respective owners.