I tried the new ChatGPT Desktop App for Linux – but I’ll stick to my browser for now

Follow ZDNET: Add us as a preferred source on Google.

ZDNET’s key takeaways

• OpenAI’s ChatGPT Desktop App for Linux is an Electron-based program.
• The application is proprietary, not open source.
• Frankly, it’s not that impressive. I’ll keep using ChatGPT in the browser.

OpenAI has released a preview of its ChatGPT desktop application for Linux. With this, OpenAI now has a native client on all major desktop operating systems and is placing its coding agent, Codex, alongside ChatGPT and its Work environment.

That’s the good news. The bad news is that, for me at least, ChatGPT continues to be annoyingly inaccurate. While I can cosplay a coder, with or without AI, my real use for AI is as an advanced search engine. When I asked the ChatGPT app on my main Linux Mint desktop to give me the technical specifications for SpaceX and Nvidia’s forthcoming Starmind AI datacenter satellites — and provided it with a link to the satellite’s site — I expected an accurate answer. I didn’t get one. Starmind’s height and wing-width, according to SpaceX, are 30m and 75m, respectively; not, as ChatGPT would have it, 20m and 70m.

Also: Microsoft is merging Copilot and Copilot 365 into one unified app – and retiring 3 features

(Disclosure: Ziff Davis, ZDNET’s parent company, filed an April 2025 lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)

Supported Linux distros 

Leaving aside ChatGPT’s lack of accuracy, this release supports Ubuntu 24.04 LTS and 26.04 LTS, Debian 13, and Fedora 43 and 44 on both x86-64 and Arm64 machines. OpenAI is distributing the client as DEB and RPM packages. In practice, this means you can run on most Debian and Red Hat-family Linux distributions, such as my own Linux Mint. 

The release also arrives roughly a month after Anthropic introduced a Linux desktop version of Claude. Anthropic’s client supports Ubuntu 22.04 or later and Debian 12 or later, giving it a somewhat broader stated baseline among Debian-derived systems. 

Also: I tested the new Claude Desktop on Linux – here’s how it compares to rival apps

Arch Linux, NixOS, and other distros are left in the lurch for now. That’s made some would-be users very unhappy. Others want the program to be distributed in Flatpak or AppImage formats. 

Still, according to OpenAI, since “Linux has been one of the most-requested platforms for the desktop app, and this launch extends ChatGPT and Codex across every major desktop operating system,” I imagine we’ll see other distros supported soon. 

A developer-oriented desktop client

The Linux application consolidates three OpenAI products that have increasingly overlapped. These are ChatGPT for general-purpose AI assistance, ChatGPT Work for delegated workplace tasks, and Codex for software development. According to OpenAI’s Linux preview announcement, the desktop app is intended to serve as a workspace for projects, files, browser workflows, and Codex sessions. For Linux developers, Codex is the headline feature.

That positioning matters in Linux, where ChatGPT has been widely available in a browser but where developers have relied on running applications in the shell, API integrations, unofficial Electron wrappers, and editor extensions for more native workflows. The new client could give OpenAI a more direct foothold on developer desktops, as AI coding tools become more tightly integrated with local projects and development environments.

Also: 75% of developers I surveyed prefer Claude Code – here’s why they choose it over Codex

For programmers, the headline features start with the ability to code local projects: Codex can examine and modify files, run commands, and work within selected project directories, with behavior constrained by your chosen permission modes. For a local project with multiple folders, the desktop app can show repositories and changed lines together for consolidated review.

The wider desktop app feature set also includes inline diff editing and pull-request review. Feature availability may vary in the Linux preview, so OpenAI has not published a detailed Linux-versus-MacOS/Windows parity table. For the fullest-featured version of the application, you still need to use a Mac.

Finally, while the app can use browser workflows, desktop integrations where supported, and plugins, each connected service’s access request must be reviewed and approved separately.

In addition, despite running on Linux, the client itself is proprietary and closed-source. That will make it a non-starter for many open-source developers. 

Also: Claude Code vs. Codex: Why I use both and how to choose one (if you must)

At launch, the Linux client does not, by default:

• Automatically scan all local disks, folders, or applications.
• Control arbitrary Linux GUI applications through native computer use.
• Access an online service merely because you have an account there. You must give it an authorized browser connection, plugin/integration, uploaded data, or an explicit task/action.
• Gain root or administrator authority just because it can run commands in a selected coding workspace.

Integration wins

I’m not too trusting of letting any AI program access my local and personal programs. However, if you’re OK with that, go ahead and integrate ChatGPT Desktop App with Google Calendar, Gmail, Outlook, Teams, Slack, and Google Drive. Once that’s done, you can then have it check your connected messages, email, calendar, documents, and project trackers. You can also pull data and tasks from your Claude workflow.

Armed with that information, it can track changes, prepare updates, and draft replies. You can also use it to create or coordinate a separate task for a specific project. 

Also: Why software factories are back – and how they work in the age of AI

Keep in mind, however, if you do that, ChatGPT gets access to all your exposed data. Specifically, anything you type, paste, upload, or deliberately add as context is sent to ChatGPT for processing. By default on consumer accounts, conversations will  be used to improve OpenAI’s models unless you turn off Settings > Data Controls > “Improve the model for everyone.” That setting applies account-wide. Chats remain in your history but are not used for model training, OpenAI says.

All that is reason enough for me to avoid this kind of functionality no matter how useful it may sound.

Oh, and for now anyway, OpenAI says native “computer use” — controlling other Linux desktop applications — is currently unavailable in this preview. That’s fine by me. It does, however, give you access to a variety of other services via plugins, such as Figma, GitHub, and Adobe. 

Sensible Linux setup

If you’re going to use the desktop app for your personal or important data, I recommend taking the following steps to use it as safely as possible. 

• Keep “Improve the model for everyone” off, unless you want chats used for model improvement.
• Use temporary chats for especially sensitive one-off queries, recognizing the stated 30-day deletion window.
• Give it a dedicated, narrow project directory. Do NOT give it access to  ~/, your entire documents tree, or a directory holding credentials and source files.
• Leave Codex on Ask for approval. Do not enable Full access unless a task truly requires it.
• Use a separate browser profile with only the service accounts required for an agent task. Avoid running it while password managers, private email, banking, or confidential source material are open.

I might add that you should take these steps regardless of what operating system you’re using. Sure, it’s easier to use AI without taking precautions, but it’s not safe. 

Trouble in Linux ChatGPT land

This new app still has a long way to go. For example, I found that my ChatGPT shell jobs wouldn’t show up on the desktop app’s project list. That’s a pretty major stopper right there. It wasn’t just me. I saw others reporting the same problem. You could see terminal sessions as standalone chats, but even when the working directory and Git branch were preserved, they really weren’t integrated. 

I also discovered that even when sitting idle with no work to do, the ChatGPT app used over a gigabyte of RAM. Adding insult to injury, I discovered the program leaked memory. After only a few hours of very light work, my PC, with 16 GB of RAM, was running out of memory. Once my computer started calling on virtual memory, it slowed to an unusable crawl. 

Also: How to create flyers with AI that people won’t hate: 8 tips (and prompts) that really work

That’s not too surprising since. After a little digging, I found it was an Electron app. These are desktop software programs written with web technologies like HTML, CSS, and JavaScript. They combine the Chromium browser engine for the user interface and Node.js for system access, allowing a single codebase to run on Windows, MacOS, and Linux. That’s fine as far as it goes, but Electron programs tend to be terribly slow and memory-inefficient compared to native operating system applications. 

Early Wayland issues emerge

I haven’t seen it myself, but initial community feedback points to the familiar challenges of shipping a desktop application across Linux’s varied display servers.

The Wayland window manager is a particular pain point. One Fedora 44 KDE Plasma user reported that Japanese input using Fcitx 5 did not work in the ChatGPT message composer when launching the application normally. The user said the issue could be addressed by starting the app with Chromium/Wayland flags, including `–ozone-platform=wayland` and `–enable-wayland-ime`. This enables the client to remain on Wayland rather than falling back to XWayland. Other forum participants reported similar behavior with Korean Fcitx5 input and raised concerns about X11 fallback and scaling on high-resolution Wayland desktops. 

Also: You can use 70+ Adobe tools without leaving ChatGPT now – here’s how

You can’t say ChatGPT didn’t warn you. The application’s page now says, “Native Wayland support is experimental and will continue to improve. In a Wayland session, the app uses XWayland when available. To explicitly select native Wayland, fully quit the app and launch it from a terminal.”

All this underscores the difference between supporting a few distributions on paper and integrating cleanly with Linux’s wide range of compositors, input methods, and workflows.

Shared problems

Thinking of workflows, my ZDNET colleague David Gewritz recently pointed out that the latest version of ChatGPT Desktop on the Mac is now missing two features he found very useful. These are the ability to take screenshots from within the app and immediately drop them into chat, and “Work with,” which lets you feed ChatGPT the contents of your current window from apps such as Notion, Notes, and TextEdit. They’re not in the Linux version either.

Also: Token-maxing is an AI cost sink – how to use agents without busting your budget

Behind the scenes, Gewritz also found that you’re often channeled into higher-token-priced service levels. You don’t have to pay for these, but if you don’t, your performance will slow down. The day of cheap AI is done. 

Strategic significance

The Linux client gives OpenAI a native route to the Linux developers who create the bulk of AI infrastructure and AI programs. The app’s central feature is not merely a ChatGPT window detached from a browser. It’s the integration of Codex and work-oriented agent features into a desktop workspace.

Honestly, I haven’t been impressed. I’m sure it will get better, although for a company with OpenAI’s billions, I’m surprised they came up with such a half-baked Electron application. Surely, they could do better. In the meantime, when I use ChatGPT, I’ll keep using it with my browser.