Tackle Codex

An OpenAI Codex provider for Laravel Tackle — run every agent on OpenAI's Codex models, with your ChatGPT subscription or an API key.

AI_CODE_PROVIDER=codex

That's the whole switch. Every Tackle agent — ai:code, ai:run, ai:fix, ai:review, the self-healer — now runs on OpenAI's Codex models. If you're signed in to the Codex CLI, usage runs on your ChatGPT plan and records as $0 in Tackle's budget; otherwise it uses your OpenAI API key with metered billing.

This is a laravel/ai provider under the hood, so it also works in any laravel/ai app without Tackle.

Installation

The easiest path from a Tackle app:

php artisan tackle:install codex

Or via Composer directly:

composer require jordandalton/tackle-codex
php artisan vendor:publish --tag=tackle-codex-config   # optional

The service provider registers a codex driver with laravel/ai automatically — no config/ai.php edits required (add a 'codex' => ['driver' => 'codex'] entry if you prefer it explicit).

Authentication

Two modes, resolved by CODEX_AUTH (default auto):

Mode	What it does
chatgpt	Uses the sign-in-with-ChatGPT tokens from the Codex CLI's auth.json. Requests go to the ChatGPT Codex backend and are covered by your ChatGPT plan — no per-token billing.
api-key	Uses OPENAI_API_KEY (or CODEX_API_KEY) against the public OpenAI API, like the stock openai provider pointed at Codex models.
auto	chatgpt when a usable auth file exists, otherwise api-key.

ChatGPT mode (subscription)

Log in once with the Codex CLI:

codex login

Tackle Codex reads (and refreshes) the same ~/.codex/auth.json the CLI uses, so one login serves both. Check what will happen before your first session:

php artisan codex:status

  Mode ............................ chatgpt (plan-covered)
  Default model ................... gpt-5.6-terra
  Auth file ....................... /Users/you/.codex/auth.json
  Account ......................... a1b2c3d4…
  Access token .................... valid, expires 2026-08-16 22:14 UTC

Access tokens are refreshed automatically shortly before expiry and written back to the auth file. The auth file location honours $CODEX_HOME and can be overridden with CODEX_AUTH_FILE.

API-key mode

CODEX_AUTH=api-key
OPENAI_API_KEY=sk-...

Configuration

Everything is env-tunable; publish the config for the full list:

Variable	Default	Description
CODEX_AUTH	auto	auto | chatgpt | api-key
CODEX_AUTH_FILE	~/.codex/auth.json	Codex CLI auth file location
CODEX_API_KEY	OPENAI_API_KEY	Key for api-key mode
CODEX_MODEL	gpt-5.6-terra	Default model
CODEX_BASE_URL	ChatGPT Codex backend	Endpoint for chatgpt mode
CODEX_API_BASE_URL	https://api.openai.com/v1	Endpoint for api-key mode

Pick the model per session as usual:

php artisan ai:code --provider=codex
php artisan ai:run "..." --provider=codex --model=gpt-5.6-terra

Tackle integration

When chatgpt mode is active, the Codex models are registered in Tackle's pricing catalog at $0/MTok — the budget tracker stays accurate (your plan covers usage) and the /model picker knows the models. In api-key mode, pin real rates in config/tackle.php under pricing.models so budget enforcement stays meaningful.

Tackle's own AI_CODE_MODEL default is a Claude model; when the provider is codex and the configured model is a leftover claude-* default, it is swapped for the Codex default automatically — AI_CODE_PROVIDER=codex really is the only env var you need.

How it works

laravel/ai's OpenAI gateway already speaks the Responses API. This package extends it with a CodexProvider + CodexGateway pair that:

• swaps the bearer token for a ChatGPT access token (with automatic refresh through the OAuth token endpoint, shared with the Codex CLI's session);
• adds the chatgpt-account-id, originator, and session headers the Codex backend expects;
• targets chatgpt.com/backend-api/codex and forces stateless requests (store: false), with encrypted reasoning carried between steps;
• serves non-streaming calls by draining a stream, since the Codex backend only streams.

In api-key mode none of that applies — it is the stock OpenAI Responses flow with Codex model defaults.

Caveats, honestly stated

Not a public, versioned API

The ChatGPT Codex backend can change endpoints, headers, or model names at any time, and plan usage is subject to ChatGPT's terms and rate limits. Every endpoint and header lives in config so a change is an .env edit, but treat chatgpt mode as best-effort. API-key mode is the stable path.

• Tackle's model-quality note applies. The agents lean heavily on tool calling; results track the model you point them at — see Models & Providers.
• laravel/ai version coupling. The gateway extends laravel/ai internals; the package pins >=0.10 <0.11 and follows the same known-risks posture as core Tackle.

Links

• tackle-codex on GitHub