OpenAI Codex Desktop Version Midnight Raid! One person commands the Agent army, programmers completely say goodbye to 996

The footsteps of Claude 5 are getting closer, and Ultraman finally can’t sit still anymore.

Just now, OpenAI unexpectedly dropped a “big move” — Codex has officially evolved into a standalone desktop app.

This is not just a window for coding; it’s an “all-in-one command center” capable of simultaneously directing thousands of agents.

Codex’s positioning is very clear: to serve as the “command hub” for agents.

Specifically, Codex can do the following:

• Seamless multitasking and switching: Call multiple AI agents to work simultaneously, isolating changes through “worktrees” so they don’t interfere with each other;

• Create and invoke Skills: Encapsulate tools and development standards into reusable capabilities;

• Set up automation workflows: Use background scheduled workflows to handle repetitive tasks automatically with Codex.

For example, if you want to add a “drag-and-drop” feature to photos in your album, selecting a “worktree” allows AI to perform its tasks within the same repository.

Codex’s evolution is chilling — it not only generates code but has also learned to use code as “Skills” to control the computer.

For instance, to fix comments in a project, you can directly invoke installed Skills, and Codex will immediately solve the problem.

Moreover, OpenAI single-handedly consumed 7 million tokens to craft a 3D racing game using just a single sentence.

This time, the birth of Codex is not just a rehash of old ideas nor a superficial “shell” packaging without sincerity.

It marks the official evolution of AI programming from a “dialogue assistant” to a “command center.”

Ultraman excitedly said, “I really love it — it’s even more surprising than I imagined!”

“AI programmers never run out of dopamine. They don’t feel frustrated, nor do they exhaust their energy. They keep going until the problem is solved.”

OpenAI President Greg strongly recommends—

I’ve been a big fan of terminals and Emacs for years, but since using Codex, returning to the terminal feels like a trip back in time — the difference in capability is obvious.

This feeling is like an AI agent interface born specifically for development.

OpenAI Codex represents a new paradigm in AI coding, very likely to reshape how developers interact with code.

Even better, Codex can team up with Claude Cowork to instantly clean up cluttered desktops.

Currently, Codex is officially available on macOS, with a Windows version coming soon.

OpenAI has also released a “limited-time benefit” — free users of ChatGPT and the GPT-4 version can now access Codex, with rates doubled for Plus, Pro, Business, Enterprise, and Edu plans.

Code-breaking weapon Codex APP debuts — One person commands all Agents

The macOS version of the Codex app is a powerful new interface.

It allows developers to easily manage multiple AI agents, process tasks in parallel, and collaborate with AI to handle time-consuming work.

In the past, the relationship between developers and AI was “pair programming” — you write a piece, it continues.

Now, the emergence of Codex will fundamentally change how software is built—

Humans no longer work closely with AI; instead, they delegate tasks directly to AI throughout the entire software lifecycle, from design and development to deployment and maintenance.

This shift has been hinted at since the release of Codex in April 2025.

The way developers collaborate with AI has undergone a fundamental change.

Existing models can handle complex, long-running tasks end-to-end, and developers are now directing multiple AI agents across projects:

Assign work, run tasks in parallel, and confidently hand over major projects that take hours, days, or even weeks to AI.

The core challenge is no longer “what AI can do,” but “how to command, supervise, and collaborate with AI at scale”—

Unfortunately, current IDEs and terminal tools are not designed for this purpose.

This new way of building, combined with enhanced model capabilities, calls for a new interaction platform.

That is why OpenAI is launching the Codex desktop app, emphasizing “a command center for a single AI agent.”

Multi-agent parallelism — code at full throttle without chaos

Codex creates a dedicated space for AI agents to work in parallel.

All AI run in independent threads organized by project, ensuring seamless task switching without losing context.

You can directly review AI’s changes within the app, comment on diffs, or even open them in an editor for manual adjustments.

It also has built-in support for Git worktrees, so multiple AI can work on the same repository without conflicts.

Each AI works on an isolated copy of your code, allowing you to explore different development paths without worrying about affecting your main codebase.

When AI agents work, you can pull (checkout) changes locally or let them continue progressing without changing your local git state.

The app automatically syncs session history and configurations from Codex CLI and IDE extensions, so you can immediately use it in your existing projects.

Unlock Skills plugins to build a 3D racing game from scratch

Codex is evolving from a code-writing AI into an AI that can truly solve problems on your computer using code.

Through Skills, you can easily extend Codex’s capabilities.

In the future, Codex will not only generate code but also handle information gathering and integration, problem-solving, writing, and more.

Skills are like packaged commands, resources, and scripts that enable Codex to reliably connect tools, run workflows, and complete tasks according to team habits.

There is a dedicated interface within the Codex app to create and manage Skills.

You can explicitly instruct Codex to use a specific Skill or let it automatically invoke Skills based on the task at hand.

OpenAI provided an example where Codex was asked to make a racing game—

With different drivers, eight maps, and even items triggered by the spacebar.

Using an image generation Skill (powered by GPT Image) and a web game development Skill, Codex independently created the game based on a single user prompt, consuming over 7 million Tokens.

It played multiple roles — designer, game developer, QA tester — verifying the results through actual gameplay.

• 60,000 Tokens

In this version, which only consumed 60,000 tokens, the graphics are quite rough.

A narrow track filled with “obstacles” that would glitch upon collision.

Skills boxes can be eaten or shot, but seem ineffective.

Most awkwardly, it loops endlessly in the “second lap”…

• 800,000 Tokens

In the 800,000-token version, the graphics improved somewhat, the track widened, and it more closely resembles typical racing games.

However, the skills obtained from boxes seem useless; after firing, cars still run separately…

And it still falls into an endless loop in the second lap.

• 7,000,000 Tokens

Finally, the 7 million token version shows much better graphics. The track is clearer, and the skill boxes are more refined.

This time, the skill boxes actually work.

At race start, we took a big move from AI without dodging. As a result, we finished last.

But compared to the previous two versions stuck in infinite loops, at least this one finished the race.

From performance testing and model training to drafting documents and reporting growth experiments, OpenAI internally built hundreds of Skills to confidently delegate previously hard-to-define tasks to Codex.

The Codex app includes a Skills library covering popular tools and workflows used internally at OpenAI, with some key examples:

• Design implementation: Pull design context, resources, and screenshots from Figma, converting them into production-level UI code with 1:1 visual fidelity.

• Project management: Handle bug classification, track releases, and manage team workloads in Linear to push project progress.

• Deploy to cloud: Use Codex to deploy your web app to popular cloud hosts like Cloudflare, Netlify, Render, and Vercel.

• Generate images: Use GPT Image-powered Skills to create and edit images for websites, UI prototypes, product illustrations, and game assets.

• Build with OpenAI API: Refer to the latest documentation when developing with OpenAI API.

• Create documents: Skills for reading, creating, and editing PDFs, spreadsheets, and files with professional layout and formatting.

Update websites using Vercel and image generation Skills

Create spreadsheets for shopping lists with spreadsheet Skills

Manage your Issue Backlog with Linear

When you create a new Skill in the app, Codex can use it anywhere you work: within the app, CLI, or IDE extensions.

You can also submit Skills to your code repository for the entire team to use.

OpenAI’s Agent Skills:

One-click automation — work for you 24/7

Codex can set up Automations to work automatically in the background on a schedule.

Automations combine commands with optional Skills and run according to your specified timetable.

When an Automation completes, the results enter a review queue, where you can review and continue subsequent work as needed.

Set up automations to create new Skills periodically

At OpenAI, teams have been using Automations to handle repetitive but important tasks, such as daily issue classification, identifying and summarizing CI failures, generating daily release briefs, and bug checks.

Dual-personality mode — switch instantly

Developers have different preferences when working with AI.

Some prefer straightforward, execution-focused partners; others like more interaction and dialogue.

Codex now allows developers to choose between two personalities — one is concise and pragmatic, the other more conversational and empathetic.

Both have the same capabilities, tailored to your preference. Just enter /personality in the app, CLI, or IDE extension to switch.

Default safety, configurable as needed

Additionally, OpenAI has integrated the “Security by Design” philosophy into the entire AI agent stack of Codex.

The Codex app uses a native, open-source, configurable system-level sandbox, just like in the Codex CLI.

By default, the AI agents can only edit files within their current folder or branch and use cached web searches.

For commands requiring higher permissions, like internet access, they will request your approval first.

You can set rules for projects or teams to allow specific commands to run with elevated permissions automatically.

Everything under code control

Today, enterprises and developers increasingly rely on Codex for end-to-end development.

Since the release of GPT-5.2-Codex in mid-December, total usage of Codex has doubled, with over 1 million developers using it in the past month.

Next, the team plans to expand Codex use cases, including launching Windows applications, pushing model capabilities further, and achieving faster inference speeds.

OpenAI scientists note that in recent weeks, they’ve written more code than in the past few years, and have used Codex to fix multiple bugs and update features in Prism.

Within the app, OpenAI will continue refining multi-AI workflows based on user feedback, making parallel task management and switching easier without losing context.

They are also developing cloud-based triggers for Automations, so Codex can run continuously in the background — not just when your computer is on.

Codex is built on a simple premise: everything is controlled by code.

The more powerful an AI agent’s reasoning and code generation, the stronger its ability across various technical and knowledge-based tasks.

OpenAI Ecosystem

However, a key challenge today is the gap between the capabilities of cutting-edge models and their ease of use in real-world scenarios.

Codex aims to bridge this gap, making it easier for people to command, supervise, and fully leverage OpenAI models in practical work.

OpenAI is committed to making Codex the most powerful AI programmer, laying the foundation for it to become a versatile AI capable of handling a wide range of knowledge tasks beyond just code.

Appendix

When creating the racing game above, the initial prompt used in Codex was as follows (condensed version):

Implement a voxel-based 3D kart racing game called Voxel Velocity using Three.js, with only one mode: single-player race (always 3 laps, 1 human vs 7 CPUs, all 8 tracks available immediately in this mode, no progress limit). Build a minimal pre-race flow including: tracks (8), characters (8), difficulty (easy/standard/harsh), optional mirror mode, optional cloning, start race, plus a menu and pause menu (resume/restart/exit). Create an arcade-style driving model with responsive controls, forgiving wall collisions, meaningful drifting as main skill, and a drift charge system producing precise acceleration levels (level 1: 0.7s, level 2: 1.1s, level 3: 1.5s), maintaining a “fast but readable” baseline speed, and continuous overtaking on wide roads. Implement exactly 8 power-ups, single-item capacity, subtle positional weighting, and mild effects (max loss of control ≤1.2s, max steering disable ≤0.6s), creating chaotic fun without hard dizziness, plus 50% reduced off-road slowdown during acceleration. Define 8 characters with stats and AI tendencies, preset CPU difficulties, track scripting with racing/spline variations, drifting zones, and obstacle avoidance, enabling AI to perform clean overtakes using multi-lane width, and publish HUD/audio elements (position, laps/last lap banner, mini-map, power-up slots, timer/segments, readable sound effects, and a music loop per track).

Subsequently, Codex was repeatedly prompted from a list of 10 general prompts to continue working on this problem.

One example prompt was:

Your task is to add new features to make the game closer to the original. First, play the game and identify what is missing compared to the original. Then select a few missing features and implement them. After each feature, thoroughly test it by playing and confirming it works correctly. If you notice any bugs while playing, prioritize fixing them.

Source: Xinzhiyuan

Risk Disclaimer and Terms of Use

Market risks are present; invest cautiously. This article does not constitute personal investment advice and does not consider individual users’ specific investment goals, financial situations, or needs. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Invest at your own risk.