Codex / ChatGPT coding with ChatGPT, Claude & SuperGrok

Last updated: 2026-08-28

Use Codex CLI and ChatGPT desktop coding / Codex mode with the same local gateway as Claude Code: ChatGPT Plus/Pro (Codex OAuth), plus Claude and SuperGrok when those logins are present.

The ChatGPT chat UI often ignores a custom provider (product limitation). The supported path is anything that reads ~/.codex/config.toml (Codex CLI, many ChatGPT coding agent surfaces).

  Codex CLI / ChatGPT coding
       │  OpenAI-compatible  ~/.codex/config.toml
       ▼
  llm-gateway  https://127.0.0.1:8787/v1
       ├── ChatGPT subscription (Codex OAuth)
       ├── SuperGrok (device / Grok CLI OAuth)
       └── Claude subscription (translated)

Related: Claude app · Claude Code · Cursor · Any app · OAuth

Prerequisites

llm-gateway helpers install
eval "$(llm-gateway helpers source)"

llm-gateway auth login chatgpt    # ChatGPT Plus/Pro via Codex OAuth
# optional: llm-gateway auth import grok
# optional: llm-gateway auth login claude
llm-gateway auth status

export KEY=local-dev
cc-gateway-up                     # https://127.0.0.1:8787

Prefer mkcert -install so Node/Electron trusts local HTTPS. Cert path after helpers install:

~/.config/inja-gateway/certs/localhost.pem

KEY=local-dev is the edge credential Codex sends to the gateway. Subscription tokens stay in the gateway auth store (~/Library/Application Support/inja-gateway/credentials.json on macOS).

ToS: personal accounts you own only.

One-command write

export KEY=local-dev
apps-write-codex          # snapshots ~/.codex/config.toml, then writes gateway provider
# or: apps-use-gateway    # Claude settings + Desktop + Codex together

Then in this shell (and your login shell / launchd if you start Codex from the Dock):

export INJA_GATEWAY_KEY=local-dev
export NODE_EXTRA_CA_CERTS="$HOME/.config/inja-gateway/certs/localhost.pem"
codex

Rollback:

apps-use-default

~/.codex/config.toml

Template: examples/apps/codex/config.toml.

model = "gpt"
model_provider = "inja"
openai_base_url = "https://127.0.0.1:8787/v1"

[model_providers.inja]
name = "Inja LLM Gateway"
base_url = "https://127.0.0.1:8787/v1"
env_key = "INJA_GATEWAY_KEY"
wire_api = "chat_completions"
# Some Codex agent builds prefer Responses:
# wire_api = "responses"

Do not put a ChatGPT Platform sk- here. The gateway holds Codex OAuth.

Models

Gateway aliases (subscription YAML, Aug 2026):

codex --model grok-4.6
# or edit model = "grok-4.6" in config.toml

List what this gateway will route:

curl -sk -H "Authorization: Bearer local-dev" \
  'https://127.0.0.1:8787/v1/models' | jq -r '.data[].id' | sort

Grok 4.6 xhigh effort is reasoning_effort on the OpenAI wire. If your Codex build exposes a reasoning-effort setting, set it to xhigh and use model = "grok-4.6" (grok-4.5 coerces xhigh → high).

ChatGPT desktop coding

Many ChatGPT desktop coding / Codex modes share ~/.codex/config.toml. After apps-write-codex:

1. Export INJA_GATEWAY_KEY and NODE_EXTRA_CA_CERTS for processes launched from the Dock (macOS GUI apps do not inherit your zshrc). A LaunchAgent, launchctl setenv, or starting ChatGPT from a terminal that already exported the vars all work.
2. Fully quit ChatGPT / Codex and reopen.
3. Confirm cc-gateway-logs -f shows path=/v1/chat/completions or /v1/responses when you send a prompt.

If the chat tab still talks to OpenAI directly, that surface has no custom-base override — use Codex CLI, Cursor, or the coding agent UI instead.

TLS

Helpers: cc-gateway-up generates certs via scripts/gen-localhost-tls.sh in the install dir.

Troubleshooting

cc-gateway-status
llm-gateway auth status
apps-codex
curl -sk https://127.0.0.1:8787/healthz
cc-gateway-logs --usage

Related

• In-repo: docs/codex-subscriptions.md
• Template: examples/apps/codex/config.toml
• Helpers: apps-write-codex · apps-codex · apps-use-gateway
• External: Codex config · Codex / ChatGPT auth