OpenAI's Codex Desktop App Turns You Into a Multi-Agent Supervisor

OpenAI shipped the Codex desktop app for macOS on February 2, 2026, and followed up with Windows support on March 4. It is not a new IDE and it is not a chat window. It is a dedicated interface for running multiple AI coding agents at the same time, across multiple projects, with proper Git integration baked in.

If you have been using Codex through the CLI or an IDE extension, this is the same underlying capability with a purpose-built shell around it. If you have not used Codex at all yet, this is probably the most accessible entry point so far.

The core problem it solves

Most developer tools assume you are working on one thing at a time. You open a file, write some code, close the loop. AI coding agents broke that model almost immediately, because you can now kick off a task, let an agent run for minutes or hours, and come back to review the result.

The trouble is that existing tools were not built for that workflow. You end up juggling terminal sessions, losing context when you switch tasks, and manually managing which agent is touching which part of your repo.

The Codex app organises work into threads, grouped by project. Each thread runs an agent in one of three modes: Local (working directly in your current project directory), Worktree (isolating changes in a Git worktree so multiple agents can work on the same repo without stepping on each other), or Cloud (running remotely in a configured environment). You can switch between threads without losing context, review diffs inline, leave comments for the agent to address, and commit or push changes directly from the app.

Skills: extending what Codex can do

One of the more interesting additions is the Skills system. A skill bundles instructions, resources, and scripts so Codex knows how to reliably connect to a tool or run a specific workflow. You can create your own, or use the built-in library, which includes:

Figma integration: Pull design context, assets, and screenshots directly from Figma and have Codex translate them into production UI code.
Linear integration: Triage bugs, track releases, and manage team workload without leaving the agent workflow.
Cloud deployment: Deploy to Cloudflare, Netlify, Render, or Vercel as part of the same task.
Image generation: Create and edit images for UI mockups, product visuals, or game assets using GPT Image.

The skills model matters because it starts to decouple Codex from pure code generation. You can ask it to gather information, synthesise content, coordinate across tools, and produce outputs that are not just source files. OpenAI frames this directly: everything is controlled by code, and the better an agent is at reasoning about code, the more capable it becomes across all forms of technical and knowledge work.

Automations: background work without prompting

Automations let Codex pick up routine tasks without you initiating each one. Issue triage, alert monitoring, CI/CD checks, PR reviews. You configure what you want it to watch for, and it acts on those triggers while you are focused elsewhere.

The current limitation worth knowing: automations only run while your laptop is powered on. Cloud-based automations are coming, but they are not here yet. For genuinely hands-off background work, that is a real constraint for now.

What this looks like in practice

OpenAI demonstrated the capability by asking Codex to build a racing game from a single prompt, complete with different racers, eight maps, and in-game items. Codex used more than 7 million tokens, worked independently across design, development, and QA, and validated its own work by actually playing the game. It is a deliberately ambitious demo, but it illustrates the direction of travel: agents handling full task lifecycles rather than individual edits.

Inside OpenAI itself, nearly all engineers now use Codex, up from just over half in July 2025. They are merging 70% more pull requests per week, and Codex automatically reviews almost every PR before it reaches production.

Broader adoption looks similar. Duolingo, Vanta, Cisco, and Rakuten are among the companies already using it at scale, and over a million developers used Codex in the month before this launch.

Technical bits worth knowing

A few things that will matter depending on your setup:

The app shares MCP (Model Context Protocol) settings with the CLI and IDE extension. Configure an MCP server in one place and the others pick it up automatically.
On Windows, Codex runs natively in PowerShell with a Windows sandbox, so you do not need WSL or a virtual machine.
Voice input is supported. Hold Ctrl+M while the composer is open, speak, and your prompt is transcribed and ready to send.
You can pop a conversation thread into a separate window, useful for front-end work where you want the agent thread visible alongside your browser or design preview.

What this means for you

If you are a developer already using Codex from the CLI or an IDE plugin, the app gives you a proper home for managing parallel workstreams. The diff review, inline commenting, and worktree support alone are worth the switch for anything beyond single-file edits.

If you have been curious about AI coding agents but found the CLI friction too high, this is a significantly more approachable interface. The project and thread organisation means you do not have to hold all the context in your head.

If you are a team lead or engineering manager, the automation and skills features are where the longer-term productivity story starts to take shape. Routine triage, PR reviews, and deployment steps are exactly the kind of work that adds up across a team and does not require human judgement every time.

On access: for a limited time, OpenAI is including Codex with ChatGPT Free and Go, and doubling rate limits on Plus, Pro, Business, Enterprise, and Edu plans. Those higher limits apply across the app, CLI, IDE extension, and cloud.

The Codex app is available to download now for macOS and Windows.