An unofficial patch for the macOS ChatGPT/Codex desktop app that adds per-task model-provider selection without requiring you to sign out of your ChatGPT account.

The patch:

• Adds a provider section to the model menu.
• Sends the selected provider when a new task starts.
• Maps model IDs to providers automatically.
• Keeps tasks from all configured providers visible.
• Keeps the normal ChatGPT login active for OpenAI models.

This project currently supports macOS only.

• macOS
• ChatGPT installed at /Applications/ChatGPT.app
• Python 3.9 or newer
• Node.js with npx

1. Click on the patch_chatgpt_providers.py file in the repository.
2. Open the menu in the top-right.
3. Click on Download.
4. Run the downloaded script:

python3 patch_chatgpt_providers.py

The installer closes processes belonging to the target app, creates a complete backup, patches app.asar, updates Electron's ASAR integrity metadata, and applies an ad-hoc signature.

Run python3 patch_chatgpt_providers.py --help to see alternate app, config, and backup paths.

Custom Codex providers belong in ~/.codex/config.toml. Provider IDs such as openrouter are referenced by the patch configuration later.

Example OpenRouter provider:

[model_providers.openrouter]
name = "OpenRouter"
base_url = "https://openrouter.ai/api/v1"
wire_api = "responses"

[model_providers.openrouter.auth]
command = "/usr/bin/security"
args = ["find-generic-password", "-a", "YOUR_MACOS_USERNAME", "-s", "Codex OpenRouter API Key", "-w"]
timeout_ms = 5000
refresh_interval_ms = 0

Replace YOUR_MACOS_USERNAME, then add or update the API key in macOS Keychain. The command prompts for the key instead of placing it in shell history:

security add-generic-password -U -a "$USER" -s "Codex OpenRouter API Key" -w

Codex also supports environment-variable authentication with env_key. Do not combine env_key with a [model_providers.<id>.auth] section. For all authentication methods and provider options, see the Codex custom model provider documentation.

Do not set a global model_provider if OpenAI and custom providers should coexist in the desktop app. The patch selects the provider when each new task starts.

Codex loads custom model metadata from the file configured by model_catalog_json.

1. Export the bundled catalog as a starting point:

   mkdir -p ~/.codex/model-catalogs
   codex debug models --bundled > ~/.codex/model-catalogs/custom.json

2. Add model objects to the top-level models array. Copy an existing entry with similar capabilities, then update its model ID, display name, context window, modalities, reasoning levels, and tool support.

   The model ID is the slug value. It must match the model ID expected by the provider, for example:

   {
     "slug": "moonshotai/kimi-k3",
     "display_name": "Kimi K3 (OpenRouter)",
     "description": "MoonshotAI Kimi K3 through OpenRouter."
   }

   Keep the remaining required fields from the copied entry and adjust them to the model's real capabilities. Do not advertise unsupported tools, modalities, or context-window sizes.

3. Point Codex at the catalog in ~/.codex/config.toml:

   model_catalog_json = "/Users/your-name/.codex/model-catalogs/custom.json"

4. Restart the app after changing model_catalog_json or the model catalog.

Use codex debug models to inspect the effective catalog Codex sees.

The installer creates:

~/.codex/desktop-model-providers.json

Example:

{
  "version": 1,
  "default_provider": "openai",
  "providers": [
    {
      "id": "openai",
      "label": "ChatGPT / OpenAI",
      "description": "Uses your signed-in ChatGPT account"
    },
    {
      "id": "openrouter",
      "label": "OpenRouter",
      "description": "Uses [model_providers.openrouter] from config.toml"
    }
  ],
  "model_providers": {
    "moonshotai/kimi-k3": "openrouter",
    "x-ai/grok-4.5": "openrouter",
    "anthropic/claude-fable-5": "openrouter"
  }
}

• providers defines the providers displayed in the app menu.
• model_providers maps exact model slugs to provider IDs for Automatic mode.
• default_provider handles models without an explicit mapping.
• Every custom provider ID must match a [model_providers.<id>] section in config.toml.
• API keys do not belong in this JSON file.

The app reloads this file when the provider menu opens and before a new task starts. Repatching is not required after editing it.

ChatGPT updates replace the patch. Run the installer again after an update.

The installer is not tied to a fixed app version or archive hash. It patches compatible source structures and stops before modifying the installed app if an update changes the relevant code.

Backups are stored by default in:

~/Applications/ChatGPT Patch Backups/

Use this script entirely at your own risk. It modifies the installed ChatGPT application in an unofficial and unsupported way.

The author and contributors provide no warranty and accept no responsibility or liability for any problems, damage, or loss caused directly or indirectly by using this script. This includes, but is not limited to, lost or corrupted chat history or other data, an unusable or "bricked" application, account warnings or restrictions, account suspension or banning, security or privacy issues, and any other direct or consequential damage.

Create and verify your own backups before running the script.

This is an unofficial modification and is not affiliated with or supported by OpenAI.