The Linux desktop preview brings ChatGPT and Codex into the same workspace. The practical questions are which package fits the machine, which features are supported and what “local” actually describes.
Start with the supported package
The August 11 community announcement introduced the preview for Ubuntu 24.04/26.04 LTS, Debian 13 and Fedora 43/44, with x64 and ARM64 builds. The current Linux guide maps Ubuntu/Debian to .deb and Fedora to .rpm, and says installation configures the signed OpenAI update repository.
The cover makes package family and CPU architecture two separate choices. An ARM64 Fedora machine needs both the RPM family and the ARM64 build; selecting the right distribution alone is insufficient. The read-only command uname -m identifies the architecture, not the distribution.
Watch the short workspace demonstration
The 45-second clip attached to the launch post shows code and document work in the desktop interface. It is a selected demonstration from that forum post, not an independent reliability test. Playback connects to the OpenAI Community media host only when started.
Credit: demonstration attached by sps to the linked OpenAI Developer Community announcement. Feature availability remains governed by the current documentation and account.
Local files do not imply an offline model
A local project describes where files and commands are accessed. It does not mean model inference happens on the laptop. For example, an agent editing a repository on disk and running its tests locally can still depend on a hosted model service. Do not infer offline operation from a file appearing in the application’s sidebar.
The Linux guide also distinguishes browser workflows from Computer Use, which is not yet supported in the Linux preview. Native Wayland is experimental; the app uses XWayland when available. The documented native-Wayland launch option is an explicit alternative, not a universal fix for every input-method problem.
When diagnosing a problem, record the application build, desktop session, architecture and exact failing action. A package that installs correctly does not certify parity for every workflow on another operating system.
Check current Linux documentation; distinguish architecture/package choice, local files from local inference, and unsupported Computer Use; include source-hosted launch clip.