patch-codex-apikey-unlock

name	patch-codex-apikey-unlock
description	Patch ChatGPT Codex (macOS/Windows) to unlock features in API key mode. Support both ChatGPT and Codex branding, show the latest hidden models returned by app-server, expose reasoning effort levels declared by models but hidden by experiment configuration, and remove API key gates from Fast/Speed, Browser, and Computer Use.
version	4.4.0

Patch ChatGPT Codex - Unlock API Key Mode Features

Overview

Remove desktop UI gates that still depend on ChatGPT sign-in context when using API key mode.

Verified with Windows Store version 26.707.3748.0 and macOS app version 26.707.31428 (Apple Silicon). The display name and main entry point have changed to ChatGPT, but package names and directories may still use Codex. The script reads the MSIX manifest or macOS Info.plist as appropriate and retains fallback support for the old name.

Unlocked Features

#	Feature	Original restriction	Patch method
1	Latest model list	API key mode has no ChatGPT hidden-model allowlist	Show every model returned by app-server
2	Hidden reasoning effort options	API key mode has no ChatGPT experiment context	Show every option the model declares in supportedReasoningEfforts
3	Fast/Speed UI	Hidden when authMethod !== 'chatgpt'	Enable the service tier selector
4	Fast request tier	Restricted to chatgpt again before sending a request	Add apikey to the allowlist
5	Plugins	Older versions restrict non-ChatGPT sessions; newer versions support API key mode natively	Patch older versions; skip automatically on newer versions
6	Browser / Chrome	API key mode has no Statsig user context	Enable the desktop gate while retaining platform and app-server capability checks
7	Computer Use	Availability and the Node runtime are restricted by Statsig gates	Enable both gates while retaining native service authentication and TCC checks
8	Voice input/dictation	ChatGPT mode only	Expand to chatgpt || apikey
9	Usage/billing settings	ChatGPT mode only	Expand to chatgpt || apikey
10	i18n localization	API key mode has no Statsig user context	Force enablement

Prerequisites

• Python 3
• Node.js (the script pins @electron/asar@3.4.1 and @electron/fuses@1.8.0 to avoid the Node 22.12+ requirement in newer releases)

Quick Start (Recommended)

# macOS / Windows; paths are detected automatically
python3 patch.py

# Check/debug only the JS patches without repacking app.asar
python3 patch.py --assets /path/to/webview/assets

# Dry run without writing any files
python3 patch.py --dry-run

patch.py performs every step automatically: stop processes -> detect the installation type -> create an independent copy -> extract/repack ASAR -> apply JS patches -> handle Windows fuses/shortcuts or update macOS ASAR integrity and signatures.

macOS

The script checks /Applications and ~/Applications in order for ChatGPT.app / Codex.app, reads the actual executable from Info.plist, and requires a Codex runtime marker in Resources. The official app is never modified. The patched copy is written beside it, with fallback to ~/Applications if that directory is not writable. The copy is named ChatGPT-Codex-Patched.app for the new brand and Codex-Patched.app for the old brand.

One-Command Patch

python3 patch.py

# Custom installation and output paths
python3 patch.py --app /path/to/Codex.app \
  --output "$HOME/Applications/ChatGPT-Codex-Patched.app"

On completion, the script refreshes the ElectronAsarIntegrity header hash in Info.plist, applies an ad-hoc signature only to the outermost app, and strictly verifies the entire bundle. Signing removes application-group/keychain entitlements that are valid only for the OpenAI Team ID while preserving runtime permissions for JIT, audio, camera, networking, automation, and similar features. Internally bundled frameworks/helpers carrying OpenAI signatures are not rewritten. To load these components with different signatures, the outer entitlement disables library validation. This weakens dynamic-library injection protection for the outer process, so this treatment is used only for the independent copy.

Browser and Computer Use

Renderer-computed desktop availability determines whether the bundled Browser/Chrome/Computer Use plugins are enabled. When an API key session lacks ChatGPT Statsig user context, the original app records reason=statsig-disabled and removes the plugins from the runtime marketplace. The patch locates the three gates by featureName and availability fields; it does not treat numeric Statsig IDs or model names as patch semantics. The original logic still performs platform, WSL, app-server experimental feature, and plugin configuration checks. The Computer Use Node runtime gate is handled separately.

On macOS, Computer Use runs the OpenAI-signed Codex Computer Use.app with bundle ID com.openai.sky.CUAService. In System Settings, the permission entry is named Codex Computer Use. Check both "Accessibility" and "Screen & System Audio Recording" (named "Screen Recording" on older systems), rather than looking for ChatGPT. Its service, client, codex, node_repl, and Node runtime all retain their OpenAI signatures.

The outer patched copy must carry an ad-hoc signature. When the macOS responsible-process check attributes the Browser native addon to this outer process, it returns missing-code-signing-identity. The patch downgrades only this result to allowed, continues to reject untrusted-code-signing-identity and missing-socket-file-descriptor, and sets the Browser socket to owner-only mode 0600; it does not modify native .node modules. The security tradeoff is that unsigned processes running as the same user lose some isolation. The official app remains unchanged as a fallback for Appshots and the TCC identity of the original main app.

Rollback

rm -rf /Applications/ChatGPT-Codex-Patched.app
rm -rf "$HOME/Applications/ChatGPT-Codex-Patched.app"
# Old brand: rm -rf /Applications/Codex-Patched.app

The official app retains its original signature and remains directly usable.

---

Windows

Rollback

Store version (MSIX, ChatGPT brand):

# Delete the patched directory directly; the original Store version is unchanged
Remove-Item -Recurse -Force "$env:LOCALAPPDATA\Programs\ChatGPT-Codex-Patched"

Traditional installer:

cd "$env:LOCALAPPDATA\Programs\Codex\resources"
Remove-Item -Recurse -Force app -ErrorAction SilentlyContinue
if (Test-Path app.asar1) { Rename-Item app.asar1 app.asar }
if (Test-Path app.asar.bak) { Copy-Item app.asar.bak app.asar }
Write-Host "Rolled back to the original version"

One-Command Patch (Windows)

python3 patch.py

Automatically detect the Microsoft Store version (MSIX) or traditional installation. Read the MSIX entry point from AppxManifest.xml and verify the Codex resource marker. Copy the ChatGPT-branded Store version to %LOCALAPPDATA%\Programs\ChatGPT-Codex-Patched without modifying the original, and create the desktop shortcut ChatGPT Codex (Patched).

---

config.toml

Preserve the existing API provider configuration. The patch does not hard-code model IDs, reasoning effort levels, or the legacy features.enable_fast setting. Models and reasoning effort levels come from the app-server model/list response (requested with includeHidden: true), and only the options that each model actually declares in supportedReasoningEfforts are shown. Fast selection uses the current version's service_tier.

---

Troubleshooting Version Updates

After ChatGPT Codex updates, both JS filenames (hash suffixes) and variable names may change. Use the following methods to locate each patch target.

General Search Strategy

# Use --assets to check/apply only the JS patches
python3 patch.py --assets /Applications/ChatGPT-Codex-Patched.app/Contents/Resources/app/webview/assets
# Windows (Store version)
python3 patch.py --assets "$env:LOCALAPPDATA\Programs\ChatGPT-Codex-Patched\resources\app\webview\assets"

# Manual search (macOS)
cd /Applications/ChatGPT-Codex-Patched.app/Contents/Resources/app/webview/assets

1. Latest Model List

grep -rl "useHiddenModels" *.js
# In version 26.707, the target is model-list-filter-*.js.
# Original condition: if(useAllowlist?allowed.has(model.model):!model.hidden)
# New condition applies only to API key mode:
# if(authMethod===`apikey`||(useAllowlist?allowed.has(model.model):!model.hidden))

2. Hidden Reasoning Effort Options

grep -rl "enabledReasoningEfforts" *.js
# There are two filtering layers in the same model-list-filter-*.js:
# 1. Remove ultra when includeUltraReasoningEffort is false
# 2. Hide disabled levels with enabledReasoningEfforts.has(effort)
# API key mode bypasses both layers but retains validation against legal enum values.
# Do not hard-code effort levels; show each model's supportedReasoningEfforts.

The full API key mode list appears in the Advanced/Effort menu. Options such as Max and Ultra appear only when the current model actually declares support. The simplified Power slider uses fixed combinations from the official app and is not modified by this patch.

3. Fast UI / Service Tier Gate

grep -rl "isServiceTierAllowed" *.js
# In use-service-tier-settings-*.js:
# allowed=auth?.authMethod===`chatgpt`,method=auth?.authMethod??null
# Change allowed to:
# auth?.authMethod===`apikey`||auth?.authMethod===`chatgpt`

4. Fast Request Gate

grep -rl "Failed to read service tier for request" *.js
# In read-service-tier-for-request-*.js:
# if(method!==`chatgpt`)return!1
# Change to: if(method!==`chatgpt`&&method!==`apikey`)return!1

5. Plugins

grep -rl "useHiddenOpenAICuratedMarketplaces" *.js
# Version 26.707 use-plugins-*.js already contains:
# return method!==`chatgpt`&&method!==`apikey`&&...
# This means API key mode is supported natively; the script should report SKIP
# instead of applying an incorrect patch.
# On the plugin detail page, connector-unavailable text is used only to explain the reason.

If an older version still has an assignment gate in check-plugin-availability-*.js, the script continues to apply the compatible false&& patch. This UI patch does not forge the ChatGPT session identity required by /aip/connectors, so do not describe newer ChatGPT session-based connectors as natively available with an API key.

6. Browser / Chrome

grep -rl 'featureName:`browser_use`' *.js
# The same desktop availability chunk should contain both:
# featureName:`browser_use`          + isBrowserAgentGateEnabled
# featureName:`browser_use_external` + isExternalBrowserUseGateEnabled
# The script traces each availability field back to its Statsig assignment and ORs
# the original gate with authMethod === `apikey`; both hooks always execute.
# It does not depend on minified variable names or use the current Statsig ID as a locator.

# The macOS Browser native pipe fallback is in .vite/build/main-*.js beside webview:
grep -rl 'browser-use-peer-authorization.node' ../../.vite/build/main-*.js
# Allow only reason === `missing-code-signing-identity`.
# Continue to reject `untrusted-code-signing-identity` and a missing socket fd.

7. Computer Use

grep -rl 'featureName:`computer_use`' *.js
# In desktop availability, locate the Statsig assignment for isComputerUseGateEnabled.

grep -rl 'computerUseNodeRepl' *.js
# The computerUseNodeRepl field synchronized to the main process has a separate runtime gate.
# The script ORs the original gate with API key mode without bypassing computer.available.
# Other authentication modes continue to use the original gate.

During runtime verification, the marketplace should contain computer-use, and the OpenAI-signed SkyComputerUseService should create the group-container socket. The TCC entry is named Codex Computer Use. Do not bypass service authentication by modifying native binaries or using codesign --deep.

8. Voice Input

grep -rn "authMethod===.chatgpt." *.js | grep -v "!=="
# Find xxx&&yyy.authMethod===`chatgpt` in annotation/comment/editor-related files.
# Expand to xxx&&(yyy.authMethod===`chatgpt`||yyy.authMethod===`apikey`)

9. Usage Settings

grep -rn "let.*===.chatgpt." *.js
# Find let r=e===`chatgpt` in usage-settings-related files.
# Expand to let r=e===`chatgpt`||e===`apikey`

10. i18n Localization

grep -rn "enable_i18n" *.js
# Find xxx=(0,YYY.useMemo)(()=>nnn?.get(`enable_i18n`,!1),[nnn])
# Change to xxx=(0,YYY.useMemo)(()=>!0,[nnn])
# Key point: without user context, the Statsig experiment gate returns false by default.

How It Works

Operation	Reason	Location
Read the entry point from the manifest/plist	The ChatGPT display name, package identity, and executable name no longer match	Installation discovery
Repack app.asar after extraction	The Owl runtime does not fall back to loading a loose app/ directory	Resources directory
Preserve .node / native module sidecars	Electron must load native extensions from disk	app.asar.unpacked
Update ElectronAsarIntegrity	Match the repacked ASAR header hash to Info.plist	Info.plist (macOS only)
Apply an outer ad-hoc signature and strict verification	Allow the modified copy to launch while preserving internal OpenAI signatures	Final step (macOS only)
Bypass hidden-model filtering	API key mode has no ChatGPT Statsig model allowlist	model-list-filter-*
Bypass reasoning-effort filtering	API key mode has no ChatGPT experiment context, so the enabled set may hide model-declared options	model-list-filter-*
Allow request-level service tiers	Showing the Fast option alone does not change actual requests	read-service-tier-for-request-*
Enable the desktop availability gate	Bundled plugins are removed when API key mode has no Statsig user context	Desktop feature chunk
Enable the Computer Use runtime gate	A Node runtime must still be selected after availability is established	Desktop feature sync chunk
Fall back when the Browser signing identity is missing	An ad-hoc outer app prevents complete responsible-process identity resolution	.vite/build/main-* (macOS only)
Bypass Statsig experiments	API key mode has no Statsig user context, so i18n/feature experiments default to disabled	webview JS
Bypass authMethod gates	Multiple features check === 'chatgpt', excluding API key mode	webview JS