The Codex app for macOS has been released, offering a robust new interface for managing multiple agents simultaneously, executing tasks in parallel, and collaborating with agents on extended projects.
This release aims to demonstrate the capabilities of Codex to a wider audience. For a limited period, Codex is being made available to ChatGPT Free and Go users. Additionally, rate limits for Plus, Pro, Business, Enterprise, and Edu plans have been doubled. These increased limits apply across all platforms where Codex is utilized, including the app, CLI, IDE, and cloud environments.
The Codex app is transforming software development processes and expanding access to creation. It supports various workflows, from pairing with a single coding agent on targeted edits to overseeing coordinated teams of agents throughout the entire software development lifecycle, encompassing design, building, deployment, and maintenance.
The Codex app: A command center for agents
Since the launch of Codex in April 2025, developer interaction with agents has undergone a significant transformation. Current models can manage intricate, long-duration tasks from start to finish. Developers are now coordinating multiple agents across various projects, delegating responsibilities, running tasks concurrently, and entrusting agents with substantial projects that may last hours, days, or even weeks. The primary challenge has evolved from understanding agent capabilities to effectively directing, supervising, and collaborating with them at scale. Traditional IDEs and terminal-based tools are not designed to support this advanced mode of operation.
This innovative development approach, combined with enhanced model capabilities, necessitates a new type of tool. The Codex desktop app serves as this command center for agents.
Work with multiple agents in parallel
The Codex app offers a dedicated environment for multitasking with agents. Agents operate in distinct threads, organized by project, enabling users to transition between tasks smoothly without losing context. The app facilitates reviewing agent changes within the thread, commenting on differences, and opening files in an editor for manual modifications.
Integrated support for worktrees allows multiple agents to collaborate on the same repository without conflicts. Each agent utilizes an isolated copy of the code, providing flexibility to explore various approaches without affecting the main codebase. As an agent progresses, users can review changes locally or permit the agent to continue working independently without altering the local Git state.
The app automatically imports session history and configurations from the Codex CLI and IDE extension, ensuring immediate usability with existing projects.
Go beyond code generation with skills
Codex is evolving from a code-writing agent to one that leverages code to accomplish tasks on a computer. Through skills, Codex’s capabilities extend beyond mere code generation to include tasks requiring information gathering and synthesis, problem-solving, writing, and more.
Skills package instructions, resources, and scripts, enabling Codex to reliably connect with tools, execute workflows, and complete tasks according to team preferences. The Codex app features a dedicated interface for creating and managing these skills. Users can explicitly direct Codex to use specific skills or allow it to automatically select them based on the task at hand.
Codex was tasked with creating a racing game, complete with various racers, eight maps, and player-operable items using the space bar. Utilizing an image generation skill (powered by GPT Image) and a web game development skill, Codex independently built the game, processing over 7 million tokens from a single initial user prompt. It assumed the roles of designer, game developer, and QA tester, validating its work by playing the game itself.
The game, along with the prompt and skills used for its creation, is provided below. Earlier iterations can also be explored to observe how Codex refined the game over extended work periods.
The game was developed by Codex using the develop-web-game skill with this summarized prompt:
Implement Voxel Velocity as a 3D voxel kart racer using Three.js, with exactly one mode: Single Race (always 3 laps, 1 human vs 7 CPU, and all 8 tracks available immediately with no progression). Build a minimal pre-race flow with only: Track (8), Character (8), Difficulty (Chill/Standard/Mean), optional Mirror Mode, optional Allow Clones, and Start Race, plus an Options menu and an in-race pause menu (Resume / Restart / Quit). Create an arcade driving model with responsive handling, forgiving glancing wall hits, meaningful drifting as the main skill, and a drift-charge system that produces exact boost tiers (Tier 1 0.7s, Tier 2 1.1s, Tier 3 1.5s) while keeping baseline speed “fast-but-readable” and pack passing constant on wide roads. Implement exactly 8 items with one-item capacity, subtle position-weighted distribution, and mild effects (max loss of control ≤1.2s, max steering disabled ≤0.6s) that create goofy chaos without hard stuns, plus off-road slowdowns that are reduced by 50% during boosts. Define the 8 characters with their given stats and AI tendencies, implement CPU difficulty presets and track-authored racing/variation splines, drift zones, and hazard avoidance so AI uses multi-lane width for clean overtakes, and ship HUD/audio essentials (position, lap/final lap banner, minimap, item slot, timer/splits, readable SFX, and one music loop per track).
Codex was subsequently reprompted continuously from a random list of ten generic prompts to continue progress on the problem. An example prompt included:
Your job is to add new features so the game matches the original more closely. First, play the game and identify what’s missing vs. the original. Then pick a few missing features and implement them. After each feature, thoroughly test it by playing the game and confirm it works. If you notice any bugs while playing, prioritize fixing them too.
Hundreds of skills have been developed internally at OpenAI to assist multiple teams in confidently delegating work to Codex that might otherwise be challenging to define consistently. These tasks range from running evaluations and managing training runs to drafting documentation and reporting on growth experiments.
The Codex app includes a library of popular skills for various tools and workflows. The complete list is available in the open source repo.
- Implement designs: Retrieve design context, assets, and screenshots from Figma and convert them into production-ready UI code with precise visual parity.
- Manage projects: Handle bug triage, release tracking, team workload management, and other tasks within Linear to maintain project momentum.
- Deploy to the cloud: Enable Codex to deploy web applications to popular cloud hosts such as Cloudflare, Netlify, Render, and Vercel.
- Generate images: Utilize the image generation skill, powered by GPT Image, to create and edit images for use in websites, UI mockups, product visuals, and game assets.
- Build with OpenAI APIs: Access up-to-date documentation when developing with OpenAI APIs.
- Create documents: A collection of skills for reading, creating, and editing PDF, spreadsheet, and docx files with professional formatting and layouts.
Updating a website using the Vercel and image generation skills
Creating a spreadsheet to generate shopping lists using the spreadsheet skill
Managing your issue backlog with Linear
When a new skill is created within the app, Codex can utilize it across all work environments: the app, CLI, or IDE extension. Skills can also be committed to a repository to make them accessible to an entire team. Information on sharing skills using team config is available here.
Delegate repetitive work with Automations
The Codex app allows for the configuration of Automations, enabling Codex to operate in the background on a predefined schedule. Automations integrate instructions with optional skills, executing according to a user-defined timetable. Upon completion, the results are placed in a review queue, allowing users to resume work as necessary.
Automations have been utilized at OpenAI to manage repetitive but crucial tasks, such as daily issue triage, identifying and summarizing CI failures, generating daily release briefs, and conducting bug checks.
Setting up an automation to periodically create new skills
A personality that fits how you work
Developers exhibit varied preferences when interacting with an agent. Some prefer a direct, execution-oriented partner, while others favor more communicative and engaging interactions. Codex now offers developers a choice between two personalities: a concise, pragmatic style and a more conversational, empathetic one. This choice does not alter the agent’s capabilities and can be selected using the /personality command within the app, CLI, and IDE extension.
Further details on setting up and utilizing the Codex app are available in the docs.
Secure by default, configurable by design
Security is integrated by design across the entire Codex agent stack. The Codex app employs native, open-source, and configurable system-level sandboxing, mirroring the approach used in the Codex CLI. By default, Codex agents are restricted to editing files within their active folder or branch and utilizing cached web search. They then request permission to execute commands requiring elevated privileges, such as network access. Users can configure rules for their projects or teams to allow specific commands to run automatically with elevated permissions.
Availability & pricing
The Codex app is now available on macOS. Subscribers to ChatGPT Plus, Pro, Business, Enterprise, or Edu plans can access Codex across the CLI, web, IDE extension, and the app using their ChatGPT login. Usage is part of ChatGPT subscriptions, with the option to acquire additional credits if necessary.
For a limited duration, Codex will also be accessible to ChatGPT Free and Go users, encouraging broader engagement with agents. During this period, rate limits for existing Codex users on all paid plans are being doubled.
What’s next
Enterprises and developers are increasingly relying on Codex for comprehensive development. Since the introduction of GPT‑5.2-Codex in mid-December, overall Codex usage has doubled, with over a million developers utilizing it in the past month. Plans include expanding Codex’s availability and utility for developers, such as releasing the app on Windows, advancing model capabilities, and implementing faster inference.
Within the app, multi-agent workflows will continue to be refined based on user feedback, aiming to simplify parallel work management and context switching between agents. Automations are also being developed with support for cloud-based triggers, allowing Codex to operate continuously in the background, independent of whether a computer is open.
Codex operates on the fundamental principle that everything is controlled by code. An agent’s proficiency in reasoning about and generating code directly correlates with its capability across all technical and knowledge work domains. A significant challenge currently lies in bridging the gap between the potential of frontier models and their practical usability. Codex is designed to address this by facilitating easier direction, supervision, and application of advanced model intelligence to real-world tasks. The focus on establishing Codex as a premier coding agent has also prepared it to function as a powerful agent for a wide array of knowledge work tasks beyond just writing code.
