Using CLI Proxy for GPT-6 Astra | Shared Grok Conversation

can I use cli proxy in order to use gpt 6 astra inside claude code using the subscription?

Yes. That’s a real, widely used setup: a local Anthropic-compatible proxy sits between Claude Code and OpenAI’s Codex/ChatGPT backend so Claude Code keeps its UI, tools, subagents, and workflow while inference comes from GPT-6 Astra on your ChatGPT/Codex subscription, not Anthropic’s API and not pay-per-token OpenAI credits. Firecrawl

How it works

Claude Code talks Anthropic’s Messages API (/v1/messages). You point it at localhost instead of api.anthropic.com:

• ANTHROPIC_BASE_URL → the local proxy
• ANTHROPIC_AUTH_TOKEN → a dummy/local token
• ANTHROPIC_MODEL / ANTHROPIC_DEFAULT_OPUS_MODEL etc. → gpt-6-astra

The proxy logs in with ChatGPT Plus/Pro (Codex OAuth), translates each request to Codex’s Responses API, and streams the reply back as Anthropic SSE. Claude Code cannot tell the backend changed. GitHub

That is subscription usage (Codex quota), not the OpenAI API meter.

Does GPT-6 Astra work yet?

Astra shipped around Sept 3, 2026. Proxy support landed quickly:

claude-code-proxy is the cleaner “just run Claude Code on Codex models” option right now. CLIProxyAPI is the bigger multi-provider gateway.

Example launch pattern with claude-code-proxy:

claude-code-proxy codex auth login
claude-code-proxy serve

ANTHROPIC_BASE_URL=http://127.0.0.1:18765 \
ANTHROPIC_AUTH_TOKEN=unused \
ANTHROPIC_MODEL=gpt-6-astra \
ANTHROPIC_SMALL_FAST_MODEL=gpt-5.6-luna \
claude

You can also map tiers, e.g. Opus → Astra, Haiku → a cheaper Codex model.

What actually works vs what breaks

Usually fine:

• chat, tools, streaming, /model, plan/agent loops
• reasoning effort mapping (low / medium / high / xhigh / max; Astra also has an ultra level some proxies are still catching up on)
• using Claude Code as the harness while Astra does the thinking

Watch for:

• Quota is ChatGPT/Codex quota, not Claude Max. Heavy agent sessions burn it fast.
• Context window mismatch. Astra is advertised with a very large window, but Codex through these proxies often reports ~272k usable. Set CLAUDE_CODE_AUTO_COMPACT_WINDOW accordingly or Claude Code will compact too early/late. GitHub
• Feature gaps. Some Claude-only tools, prompt-cache economics, and computer-use paths will not map 1:1.
• Client fingerprint. Older CLIProxyAPI builds fail Astra even if your ChatGPT plan has it.

Terms / account risk

This is unofficial.

• Using a ChatGPT subscription via an unofficial Codex proxy is against the spirit (and often the letter) of OpenAI’s consumer terms. It can get rate-limited or the account flagged.
• Do not feed a Claude consumer OAuth token into a third-party proxy. Anthropic has blocked those tokens outside official Claude surfaces since early 2026. For this setup you want ChatGPT/Codex login on the proxy, and a dummy token on the Claude Code side. Vallettasoftware

There is also a more “official-ish” path some writeups mention: OpenAI’s Codex plugin for Claude Code. If you care about ToS, start there before a community proxy.

Short answer: yes — use a current Claude↔Codex proxy (claude-code-proxy is the most Astra-ready), log in with ChatGPT/Codex OAuth, point Claude Code at localhost, set the model to gpt-6-astra. It works in practice; treat it as unsupported and quota/ToS-sensitive.

How to work with GPT-6 Astra in terminal

Use Codex CLI. That is the Claude Code equivalent: a terminal agent, with Astra as the model.

1. Install
   • macOS / Linux:

   curl -fsSL https://chatgpt.com/codex/install.sh | sh
   

or

npm install -g @openai/codex

codex --version

You want 0.153.1+. Astra became selectable around 0.153.1 and the bundled default around 0.153.4. Blakecrosley

2. Sign in (subscription)

   cd /path/to/your/repo
   codex
   

   Choose Sign in with ChatGPT (Plus / Pro / Business / Enterprise). That uses Codex quota on the plan, not API credits.

3. Run Astra

   codex -m gpt-6-astra
   

or start codex, then:

/model

and pick gpt-6-astra.

4. Daily loop (maps to Claude Code)

Example session:

cd ~/projects/app
codex -m gpt-6-astra

Then:

Find why checkout fails on empty carts and fix it. Run the tests.

Codex reads the repo, edits files, runs commands in its sandbox, and asks when it needs extra permission.

5. Practical takeaway Claude decides and writes. Codex checks, argues, and takes the jobs Claude is bad at or too expensive for. Max 20x = conductor + implementer. OpenAI 5x = reviewer, adversary, rescuer, mill.