Saw Desktop Agent Center trending on HN this morning. Local AI, fully offline, triggered by a global hotkey — you press a combo, a panel pops up, you type what you want, and it launches an application or runs a script you wrote at some earlier point. Clean idea. The thread is full of people excited about cutting cloud latency out of small automation tasks, and I get the appeal.

But the hotkey is doing all the work in that design. And without the keystroke, the AI has no idea what you want or where you are. So that’s fine for “open Spotify” or “summarize my clipboard.” But it falls apart the moment the work is tab-specific.

The hotkey is a contextless trigger

A hotkey says wake up. But it doesn’t say what about. So the user has to fill in the rest with words: type the prompt, type the target, type whatever metadata the agent needs to do its job. For desktop tasks that have to span multiple apps, this is actually the right primitive. Because you can’t assume the agent knows which window you meant.

But browser work is the opposite. Almost every browser task is tied to a specific tab. The Gmail draft you want help with. The pricing page you’re trying to extract a table from. The LinkedIn profile you’re researching. The Stripe invoice you need to download. And none of these need a hotkey, because the page itself already encodes most of the prompt.

And once the AI can see the tab, the trigger problem mostly evaporates. So you stop typing “draft a reply to the email from Sarah about the Q2 review.” You just say “draft a reply” with the email already on screen.

Side panel beats global hotkey

This is why I think the Comet/Operator/cloud-agent crowd got the form factor wrong even when they got the model right. They built the AI as a separate browser, or a separate window, or a separate cloud session that has to spin up its own Chromium and re-login to everything you were already logged into. Hotkey-style summoning, just slower and clunkier.

But Dassi takes a different bet. It runs inside the tab you’re already on, as a Chrome side panel. So there’s no summoning. The current page is the context: your logins, your cookies, your scroll position, the half-typed message in the compose box. When you open it, it already knows what you’re looking at. And that’s not a feature so much as the consequence of living where the work is.

Where hotkeys still earn their keep

Hotkeys are great when:

  • the work spans applications (clipboard, terminal, Slack)
  • there’s no natural surface for the AI to attach to
  • you want to invoke from anywhere on the OS
  • the prompt is the whole input, not a modifier on what’s visible

Desktop Agent Center fits all four. So does Raycast AI. So does Alfred. And hotkeys win on the desktop because the desktop has no canonical place for an AI to live.

The browser has a canonical place

It’s the tab. Tabs are already the unit of work, the unit of context, the unit of authentication. So tying the agent to the tab gives you the trigger for free.

What “listens to your tab” actually means

A side-panel agent that observes the active tab gets a stack of signals a hotkey agent never sees. Which domain you’re on. Whether you’re logged in. Which form fields are already filled. What text is selected. Whether there’s a modal blocking the page. And the DOM, in other words, is a richer prompt than anything a person would bother to type in a popup, and it sits in front of the agent the whole time, refreshing whenever the page itself does.

So when I open Dassi on a Notion page and say “extract the bullets into a table,” it doesn’t ask which Notion page. It already knows. So that’s the trigger paradigm.

And this is also why cloud browser agents keep tripping over themselves. They spin up a clean Chromium that has none of your context, then burn the first ten actions logging in and dismissing cookie banners. I wrote about that gap in Cloud Browser Agents Can’t See Your Tabs, and it has not closed since. So the same way desktop agents went local to escape cloud latency, browser agents need to stay in the actual browser to escape context loss.

Different ergonomics for different work

So both trigger paradigms are right, just for different kinds of work. Hotkey AI for desktop automation. Tab-aware AI for browser automation. And trying to force one model onto the other gets you Comet (a whole new browser when a side panel would do), or a hotkey agent that can’t see the page it was summoned over.

If you spend most of your day in tabs, the side panel is the better seat. Dassi is free, lives in the Chrome side panel, and runs on your own LLM key. No global hotkey to memorize. Just the tab you were already on.