ChatGPT for Linux is now available — I tried the new OpenAI app

OpenAI has released a preview version of ChatGPT Desktop for Linux. The client supports Ubuntu, Debian, and Fedora, and integrates ChatGPT, the Work workspace, and the Codex programming agent. Now, the native OpenAI application is available for all major desktop operating systems.

The news is good, but the application has not yet convinced me to give up my browser. The reason is not only technical shortcomings: ChatGPT still provides inaccurate answers.

I primarily use AI as an advanced search engine. During testing on a computer running Linux Mint, I asked the application for specifications of the upcoming Starmind satellites from SpaceX and Nvidia for AI data processing centers and provided a link to the satellite's website. ChatGPT indicated a height of 20 m and a wingspan of 70 m. According to SpaceX, the correct values are 30 m and 75 m, respectively.

Supported Distributions

The preview version officially works on the following systems:

• Ubuntu 24.04 LTS and 26.04 LTS;
• Debian 13;
• Fedora 43 and 44.

Both x86-64 and Arm64 architectures are supported. OpenAI distributes the application in DEB and RPM packages, so it can be installed on many derived distributions of the Debian and Red Hat families, including Linux Mint.

About a month earlier, Anthropic released a desktop version of Claude for Linux. It supports Ubuntu 22.04 and newer, as well as Debian 12 and newer. Formally, Claude's requirements for Debian-based systems are broader.

Arch Linux, NixOS, and other distributions are not yet included in the list. Some users are also requesting the release of ChatGPT in Flatpak or AppImage formats. OpenAI refers to Linux as one of the most requested platforms for the desktop client, so it's reasonable to assume that the list of compatible systems will expand over time.

Codex and Local Projects

ChatGPT has been operating on Linux through the browser for a long time. Developers have also used command line tools, APIs, editor extensions, and unofficial Electron shells. The new client aims to consolidate projects, files, browser scripts, and Codex sessions in one workspace.

For developers, the central feature remains Codex. The agent can explore and modify files, execute commands, and work in selected directories. Access depends on the user-selected permission mode.

When working with a project across multiple folders, the application displays repositories and modified lines in one window. Built-in edit differences and pull request validation are also promised.

The feature set of the Linux preview version may differ from the clients for macOS and Windows. OpenAI has not published a detailed comparison table, and the most complete version of the application is currently available on Mac.

Access to each connected service must be verified and confirmed separately. The client remains proprietary, with closed source code — for many open-source supporters, this is enough to refuse to use it.

At the time of launch, ChatGPT Desktop for Linux does not support by default:

• automatically scanning all local drives, folders, and applications;
• managing arbitrary graphical applications via native computer interaction functions;
• connecting to an online service simply because the user has an account there;
• gaining root or administrator rights solely because Codex is allowed to execute commands in the selected development environment.

Access to online services requires providing authenticated connection through a browser, plugin, or integration, uploading data, or explicitly assigning the necessary action to the application.

Integrations and Data Access

ChatGPT Desktop can be linked to Google Calendar, Gmail, Outlook, Teams, Slack, and Google Drive. The application will be able to access connected messages, emails, calendars, documents, and project tracking systems. It can also receive data and tasks from the Claude workflow.

Based on this information, ChatGPT can track changes, prepare updates, draft responses, and coordinate individual tasks within a project. Plugins for Figma, GitHub, and Adobe are also available.

I am not ready to give AI such access to personal and local data. Everything the user inputs, pastes, uploads, or intentionally adds to the context is sent to ChatGPT for processing.

In consumer accounts, conversations can be used by default to improve OpenAI models. This can be disabled in the Settings > Data Controls menu using the "Improve the model for everyone" option. The setting applies to the entire account. According to OpenAI, after disabling it, chats will remain in history but will not be used for training models.

For me, this is enough to avoid such integrations, no matter how convenient they might seem.

The native "computer usage" function, which would allow controlling other Linux applications, does not work in the preview version. I find that entirely satisfactory.

How to Mitigate Risks

If you still plan to use the application for personal or important data, it is better to limit its capabilities from the outset:

• Turn off "Improve the model for everyone" if you do not want to send conversations for model improvement.
• For one-time confidential requests, use temporary chats. OpenAI states a 30-day deletion period for them.
• Allocate a separate directory for the application with clear boundaries. Do not grant access to the entire ~/, document tree, or folders with sensitive information and important source files.
• Keep Codex's mode as Ask for approval. Do not enable Full access unless necessary.
• Create a separate browser profile and only authorize the services the agent needs.
• Avoid running the application alongside open password managers, banking services, personal email, or sensitive materials.

These measures apply not only to Linux. Working with AI without restrictions is easier but unsafe.

Terminal Projects Not Integrated with the Application

During testing, I found that ChatGPT tasks launched from the command shell do not appear in the desktop client’s project list. Other users have reported the same issue.

Terminal sessions were displayed as separate chats. The application retained information about the working directory and Git branch, but there was no full integration between the environments. For a tool focused on developers, this is a significant limitation.

More than a Gigabyte of RAM at Idle

Even without active work, ChatGPT Desktop used over 1 GB of RAM. Then a memory leak appeared: after a few hours of very light load on my computer with 16 GB of RAM, available memory began to run low. The system switched to virtual memory and slowed down to an unacceptable level.

The application is built on Electron. This platform uses HTML, CSS, and JavaScript, the Chromium engine for the interface, and Node.js for system access. This allows a single codebase to run on Windows, macOS, and Linux.

This approach is common, but applications on Electron often fall short of native programs in speed and memory usage. In the case of ChatGPT Desktop, this was quite noticeable.

Wayland Currently Has Limitations

I personally did not encounter any Wayland issues, but early user feedback indicates that the application is still poorly adapted to the variety of graphical environments in Linux.

A Fedora 44 user with KDE Plasma reported that when normally launched, Japanese input through Fcitx 5 did not work in the ChatGPT message field. The issue was worked around by launching the client with Chromium/Wayland flags:

--ozone-platform=wayland--enable-wayland-ime

This allowed the application to continue working through Wayland instead of switching to XWayland. Other forum participants described similar problems with Korean input via Fcitx5. Users are also concerned about reverting to X11 and scaling on high-resolution Wayland displays.

OpenAI has already added a warning to the application's page. The company labels native Wayland support as experimental and promises to improve it. In a Wayland session, the client uses XWayland if available. To forcefully launch the application in native Wayland mode, it must be completely closed and opened from the terminal.

Formal support for multiple distributions is not enough: the client still needs to work properly with different compositors, input methods, and Linux desktop environments.

Not All Features Available Even on Mac

My colleague from ZDNET, David Gewirtz, noted that two useful features disappeared from the latest version of ChatGPT Desktop for Mac. The first allowed users to take a screenshot directly in the application and immediately send it to the chat. The second feature, Work with, transferred the contents of the current window from Notion, Notes, TextEdit, and other programs to ChatGPT. These features are also absent in the Linux client.

Gewirtz also discovered that the service often directs users to levels with higher token costs. Paying for them is not mandatory, but refusal may decrease performance. The era of cheap AI is over.

For Now, the Browser is More Convenient

The release of the Linux client is strategically important for OpenAI. The company has gained direct access to developers who create a significant portion of the infrastructure and software for AI. The primary value of the application is not in a separate ChatGPT window but in integrating Codex, local projects, and agency work functions.

However, the current preview version did not impress me. Inaccuracies in answers, weak integration with terminal sessions, high memory usage, leaks, and Wayland issues make the client feel unfinished. It is particularly strange to see an Electron application from a company with OpenAI's resources.

It will likely improve over time. For now, I will continue to use ChatGPT in the browser.