OpenAI’s Linux ChatGPT app is here, and Codex comes with it
Most of what you have heard about the new ChatGPT desktop client for Linux is about parity. Mac has had it since July, Windows has it, Linux is catching up. That framing misses the more interesting question. Linux is the first platform where the unified ChatGPT-plus-Codex app ships on day one as native packages, with x64 and ARM64 binaries, and with Work mode for the kind of multi-step coding tasks that previously required a separate tool.
If you are a developer on Linux, the change worth your attention is not the parity story. It is that you can now point an agentic coding assistant at a real local project folder, watch the diffs, and keep your existing editor. No more CLI-only sessions, no more browser tabs that lose local context, no more IDE plugin that vanishes when you switch tools.
The shift is small enough that most people will overlook it. The unified app is one binary, one install, one login. The agent lives in the same window as the chat. For the first time on Linux, the entire agentic workflow is in a single tool that does not depend on your editor of choice.
What the new Linux build actually does
The desktop app runs as a native binary with system integration, not a wrapper around a website. You install a real DEB or RPM, you get updates through your distribution channels, and you do not need to maintain a hand-rolled script to keep it current. The app uses the same login as the web version, so there is no separate account or API key to manage.
On launch, you land in the chat window most people already know. The new piece is in the upper-left menu, where you can flip into Codex without opening a second app or juggling credentials. Codex in this context is not a chatbot. It is an agent that opens a folder, edits files, runs commands, and works with Git on the project you point it at. The agent does only what you allow it to do, and the permission prompts are the part you should read carefully.
Who this is for, and who should wait
The honest answer is that the desktop client on Linux is for two audiences right now. The first is developers who have been waiting for an officially supported agentic coding environment that does not lock them into a specific IDE. The second is people who want to use ChatGPT on Linux the same way they use it on Mac or Windows, with system-level polish like a tray icon, notifications, and the kind of window management that browser tabs do not give you.
Developers who have been making the CLI version of Codex work for the last year will find the desktop app is an upgrade in ergonomics, not in capability. Casual ChatGPT users can stick with the web app on Linux, which still does the job without any install. Enterprise fleet operators should hold off until the preview stabilizes.
What I would do this weekend to evaluate it
Here is the install plan I would run on a fresh Ubuntu 24.04 box if I were trying this for the first time. It is also the order I would recommend to anyone who wants to form a real opinion of the preview instead of skimming the announcement.
- Pull the DEB from OpenAI’s download page, verify the SHA256, install with
dpkg - Log in with the same credentials as the web client
- Open a small personal project, not a work project, as your first Codex target
- Grant Codex access to the project folder only, not your home directory
- Try a docstring addition or a small refactor first, before any bigger task
This is the same progression I use when evaluating any agentic coding tool. Start with the boring task that you can verify by hand, then graduate to something harder once you trust the tool’s diffs.
Permission scoping is the part you cannot skip
Codex reads and writes files based on the permissions you grant. If you scope it loosely, the agent can do things you did not intend, including editing or deleting files outside the project you meant to give it. The same permission model that makes the tool useful for legitimate work also makes it dangerous if you point it at the wrong directory.
The pattern I would adopt from day one is to treat the permission prompt the same way you treat sudo. Read what it asks for, scope the access to the specific folder you are working in, and never blanket-approve across multiple projects at once. On a work machine with proprietary code, check your company’s policy on sending source code to third-party AI services before you wire this in, even if the data is supposed to stay inside your own projects.
Distribution support and what is missing
The preview covers the mainstream desktop distributions cleanly. Ubuntu 24.04 LTS, 26.04 LTS, Debian 13, Fedora 43, and Fedora 44 all have official packages from OpenAI. Both x64 and ARM64 are supported from day one, which is unusual for a Linux preview. Arch users have unofficial packages in the AUR that usually catch up within a day or two of any major release.
What is missing is system-tray polish. The macOS client has more mature notifications, a proper tray icon behavior, and window-management quirks that the Linux build has not caught up to yet. Most of the core features work, but you will probably hit one moment where the app behaves slightly differently than on a Mac. That is normal for a preview, and the same lag exists on the Windows client.
Trade-offs
Preview-grade stability is the first trade-off worth naming. Bugs are expected, and the official line is that the app is not production-ready. If your team relies on ChatGPT for daily work, do not make this your only interface. Keep the web app open in a tab as a fallback while you evaluate.
Distribution coverage is the second trade-off. The official package list covers Ubuntu, Debian, and Fedora. If you run Arch, NixOS, or anything more exotic, you are depending on community packages. Most of the time that is fine, but the day a critical bug lands, you will be waiting on a maintainer instead of OpenAI.
The agent permission model is the third trade-off. The whole point of Codex is that it can edit files on your machine, which is exactly what makes it dangerous if you scope permissions loosely. The preview gives you the controls to scope it tightly, but you have to actually use them.
Bottom line
If you develop on Linux and you have been waiting for an official agentic coding tool, the wait is over. Install the package for your distro, log in, and start with a small personal project. Keep permissions scoped, use Work mode for tasks you would otherwise spend an afternoon on, and check the changelog after each update. The parity framing sells the change short. The real story is that Linux developers now have a first-party option for the kind of agentic coding sessions that used to require juggling tools, and the option is mature enough to evaluate on a real project.