The Codex Changelog
All Episodes
ChatGPT Lands on Linux with Native Codex and Easy Imports

ChatGPT Lands on Linux with Native Codex and Easy Imports

0:00|0:00

The hosts break down ChatGPT’s native Linux desktop preview, including supported Ubuntu, Debian, and Fedora builds, local Codex execution, and one-command imports from Claude Code, Claude Cowork, and Cursor. They also dig into Wayland input bugs, CLI-to-desktop project mapping quirks, and the migration pressure caused by model sunsets.

Show Notes


Chapter 1

Native Linux Desktop Arrives with Cross Agent Imports

Ethan Park

So I, I was actually looking at my terminal history yesterday, and I, I realized I haven't opened WSL or spun up a background virtual machine in... man, four days now. Because the preview is, it is finally here. The ChatGPT desktop app for Linux is available in preview for supported Ubuntu, Debian, and Fedora desktop distributions. And it runs native Codex execution right on the metal.

Maya

Wait, native? So no subtle translation layer, no running inside a heavy web view wrapper pretending to be a system app?

Ethan Park

None of that. Native deb and rpm packages built specifically for x64 and ARM64 architectures. We are talking full official builds for Ubuntu 24.04 LTS and 26.04 LTS, Debian 13, and Fedora 43 and 44. Oh, and by the way, before we dive into the process architecture, a quick shout out to Jellypod for making our daily show possible. But yeah, Maya, no virtual machines needed.

Maya

Okay, Ubuntu 24.04 LTS and 26.04 LTS getting direct package love right out of the gate is massive. But, okay, let me stop you there, Ethan, because the thing that caught my eye wasn't just the execution loop. It was the cross agent sync engine. You can literally bring over your entire environment from Claude Code, Claude Cowork, or Cursor with one command.

Ethan Park

Wait, how seamless is that import actually? Is it just pulling basic system prompts, or is it grabbing full workspace context?

Maya

It grabs everything. You go to Settings, then Import, or you just type slash import in the CLI, and it ingests custom skills, settings, background plugins, prompt instructions, and active workspace state. So if you spent weeks fine tuning skills inside Cursor or Claude Code, Codex ingests them in the background with auto updates so you aren't re-keying configurations.

Ethan Park

Hmm, that auto update piece is big because workspace drift is usually where multi tool workflows go to die. And on the local execution side, they implemented native Linux process sandboxing. So when Codex runs code or spins up a local dev server, it isolates those process trees locally, while letting the bundled browser plugin render file backed pages right inside the app frame.

Maya

Which sounds incredible until you look at how OpenAI actually decided to distribute the thing! I mean, deb and rpm packages for Debian 13 and Fedora 44 are great if you stay inside those standard distributions. But why on earth did they skip Flatpak or AppImage for an initial Linux preview?

Ethan Park

Ah, you are thinking about the Arch and NixOS folks, aren't you?

Maya

Yes! If you are on Arch Linux or NixOS, you are literally forced to pull out utilities like alien or bsdtar just to manually extract the binary payload from the deb file and patch your library paths manually. It creates this weird unnecessary friction for the exact power users who want local agent execution most.

Ethan Park

That is a very fair point. You get raw native performance on Ubuntu 26.04 LTS, but if you step two inches outside the main distribution path, you are unpacking archives by hand.

Chapter 2

Wayland Input Pitfalls, CLI Mapping Bugs, and Model Sunsets

Maya

And speaking of edge cases on Linux, we need to talk about Wayland. Because if you use Wayland compositors like Hyprland, there is a very specific input bug waiting to jump scare you.

Ethan Park

Wait, is this the virtual keyboard issue? I saw someone talking about speech to text macros getting completely garbled.

Maya

Yes! If you use automated macro tools or speech to text like wtype to inject text into the ChatGPT app under Wayland, all your letters turn into random numbers and symbols. It is like the virtual keyboard map completely loses its mind.

Ethan Park

Oh, weird. So what is the workaround? Is it a permission flag?

Maya

No, you have to ditch wtype for that specific workflow and route through ydotool using uinput at the system level instead. Once you route through uinput, the virtual keycodes map cleanly into the desktop window again.

Ethan Park

That is a super handy tip. And speaking of state confusion, there is another subtle gotcha with the CLI to Desktop project mapping. If you have been using the Codex CLI on your machine and then open the brand new Desktop app, your existing CLI sessions do not automatically group into Desktop project folders.

Maya

Wait, they don't sync up? So where do those active CLI terminal sessions go?

Ethan Park

They show up as standalone unlinked chats floating in your main history thread. To stop your state from fragmenting, you have to launch the desktop app directly from your project's working directory or explicitly bind the project directory inside the GUI settings before running commands.

Maya

Right, so if you don't anchor the working directory first, the app treats your terminal session like an isolated one off query instead of a project component. Good to know.

Ethan Park

Exactly. Now, while we are talking about setup shifts, there is a strict migration deadline every team needs on their calendar right now. On August 31, 2026, gpt 5.4 and gpt 5.4 mini will be retired for authenticated ChatGPT sign ins.

Maya

August 31, 2026? That is coming up fast. What are the recommended drop in replacements?

Ethan Park

You need to update your config files and environment variables to gpt 5.6 terra and gpt 5.6 luna. If your desktop or CLI tools hardcode gpt 5.4, those calls are going to fail the moment September hits.

Maya

So swap gpt 5.4 for gpt 5.6 terra, and swap gpt 5.4 mini for gpt 5.6 luna. Got it.

Ethan Park

Right. And for my practical workflow tip today: when you move those long running terminal tasks over to the desktop project view, make sure you double check your workspace trust settings. By default, native process execution will prompt for permission on network calls until you explicitly trust the workspace directory root in your config flags.

Maya

That saves about fifty unnecessary prompt popups a day. Alright, update your models to gpt 5.6 terra, point ydotool at uinput, and grab those deb packages. Talk soon, everyone.