OpenAI finally shipped a ChatGPT desktop app for Linux, and I saw it on Hacker News sandwiched between a thread about Desktop Agent Center and the recurring argument over whether any of this has produced a measurable productivity gain yet. Three stories, one pattern. Native keeps arriving. Context doesn’t.

Because a desktop app is a window. That’s the whole product. A window that no longer lives in a browser tab, with a global hotkey, a dock icon, and probably saner keyboard handling than the web version. All genuinely nice things for anyone who’s been running ChatGPT in a pinned Firefox tab for three years. None of them change what the model can see the moment you start typing.

The install is local. The context is not.

There’s a slippage in how people use the word “local” that I think does real damage to how folks evaluate these tools. Running on your machine sounds like it ought to mean knowing something about your machine. It does not. The Linux app is a client that talks to OpenAI’s servers exactly like the tab did, and its knowledge of your working life begins and ends at whatever you paste into the composer.

So the Gmail thread you’re staring at is invisible. The CRM record in the next workspace, the half-filled expense form, the Jira ticket with eight comments you wanted summarized: none of it exists to the app until you select it, copy it, alt-tab, and paste it.

Count the actual motion

I did this for a week, on purpose, keeping a tally like a bird watcher. Draft a reply to a customer email: open the thread, select the last three messages, copy, switch windows, paste, type the instruction, wait, read the output, copy it back, switch again, click into the reply box, paste, then spend forty seconds fixing the parts where the model invented a detail because it never saw the earlier messages in the thread that I hadn’t bothered to copy. That’s eleven discrete actions for a task that is, in substance, “answer this person.” Fill in a vendor form from a PDF quote: same shape, more switching. Pull the ten rows out of an admin dashboard that refuses to export without the enterprise plan: I can’t even paste that one usefully, because the table is paginated and the copy comes out as a wall of unaligned text.

The frustrating part is that the model handles every one of those tasks well once it can see the input. The intelligence isn’t the bottleneck and hasn’t been for a while. The bottleneck is that the smartest thing on my desk is sitting in a window with a wall between it and the six authenticated tabs where my actual work is stored, and I am the wall’s only door. Every keystroke of that copy-paste shuffle is me acting as a very expensive, very slow data bus. A desktop binary doesn’t remove the wall. It just gives the wall a nicer frame and a spot in my app launcher.

Same window, same session

The alternative isn’t a smarter model or a bigger context window. It’s putting the assistant inside the browser where your logins already live, which is what a side panel extension does. Dassi runs in the Chrome side panel, reads the page you’re on, and acts through the session you’re already authenticated in. No OAuth grant, no API token for Gmail, no copying the thread anywhere.

And you can point it at the model you’re already paying for. Log in with your ChatGPT subscription, or bring a key for Claude or Gemini. We wrote about that setup in more detail here. The model is the same. The plumbing is what changed.

Linux users specifically

You’ve been asking for this app for two years. Now you have it, and it can’t read the terminal, the editor, or Chrome. Which is where all your damn work is.

What I’d want instead

An honest caveat: for anything that isn’t tied to a page, the desktop app is fine and I’ll keep using one. Brainstorming, rubber-ducking a design, rewriting a paragraph. Work that starts from nothing doesn’t need context.

But most work doesn’t start from nothing. It starts from a screen someone else built, behind a login, in a tab I’ve had open since Tuesday. Desktop AI went local and browser agents mostly went the other direction, which is a weird crossing of the streams that nobody planned.

Anyway. Congrats on the .deb, I guess.