Why We're Building a Desktop App in 2026

Ask developers where they spend the working day. The 2025 Stack Overflow survey answers with a list, and the list has a property nobody remarks on:

Stack Overflow Developer Survey 2025, development environments, multi-select, so shares exceed 100% combined.

Every entry is a desktop app. The decade that moved email, docs, design, and project management into the browser left the tools developers live in untouched, and the company that owns the world's largest web platform ships its flagship editor as a local program.

We're building ICE, our Integrated Cloud Environment, as a desktop app, and in 2026 that reads as contrarian. My claim is that it's the opposite: for instruments, the tools a professional plays for hours a day, desktop never lost, and infrastructure tooling is an instrument that got misfiled as a website.

Documents versus instruments

The browser's wins share a shape: the thing being worked on is a shared artifact, and the URL is the artifact. Docs, tickets, wikis. Distribution beats latency for those, because the collaboration is the product.

Instruments have the opposite shape. An editor, a terminal, a debugger: you inhabit them, they own your keyboard, and they touch local resources constantly. A browser tab can't fully own the keyboard; Cmd-W is always one reflex away from closing your session, and the tab boundary keeps your tool at arm's length from the filesystem, the OS keychain, and your running processes. Latency compounds too. A person who triggers an interaction thousands of times a day feels every added millisecond as friction, which is a large part of why 75.9% of the industry works in a local editor while writing software for the cloud.

There's a quieter dependency too: instruments get customized. Keybindings, themes, dotfiles carried between jobs like family recipes. A tool you shape to your hands is a tool you keep, and the browser makes that shaping shallow, because in a tab the state belongs to the site rather than to you.

An operating session is a place, not a page

Cloud consoles are forms. Stateless, per-request, amnesiac: every visit starts from a dashboard that has forgotten your context, and the context is the job. Working on infrastructure means holding a model across an afternoon, and the tab-shaped version of that model evaporates on every reload. Anyone who has rebuilt a seven-tab investigation after one accidental window close knows the tax.

There's a darker version of this argument, and October 2025 supplied it. During Azure's October 29 outage, the Azure portal itself was impaired, meaning the tool for managing the blast radius was inside the blast radius. The local-first movement wrote the principle down years before that outage made it vivid:

Your model of your infrastructure, the map you need most during an outage, shouldn't be hosted inside the thing that's on fire. ICE's graph lives on your machine. The clouds can be down and the map still opens.

Credentials follow the same logic. A browser-based tool that talks to AWS, GCP, and Azure on your behalf generally means your cloud credentials transit somebody's backend, and that somebody becomes part of your attack surface. The design goal of a desktop app is blunter: keys stay in the OS keychain, API calls go from your machine to your clouds, and we never hold what we can't lose.

Figma is the counterargument

The steelman has a name. Figma beat entrenched native design tools from inside a browser tab, and the reasons were real: nothing to install, a URL is the file, multiplayer by default. Distribution through a link is a genuinely superior adoption model, and no update ever ships late to a browser. I'll concede the blurry boundary too. Plenty of "desktop apps" are Electron, a bundled browser in a native coat, so the war looks over if you squint. And building desktop means carrying the update, signing, and packaging tax that the web abolished; we pay that tax and it's not small.

But look at what Figma's users collaborate on: the canvas itself, an artifact safe to share by URL. Infrastructure's object is a live system wearing credentials, where "anyone with the link" is the beginning of an incident report, and where the collaboration layer already exists and is called git. And Electron's popularity argues my side, quietly: when teams could ship a tab, they still chose a desktop shell to get the keyboard, the tray, the keychain, and the filesystem. The UI technology surrendered. The deployment target didn't.

Defending this choice publicly is the point of making it. ICE is in development, the desktop decision is made, and if it's wrong we'll find out in the most instructive way available. My bet: the people who spend all day operating infrastructure will want what people who spend all day writing code already have, a local instrument with the full model inside it.

Here's a test you can run on yourself. The last time you were paged, how long did you spend finding the right console tab and re-authenticating before you could even look at the problem? Now compare that with how long your editor takes to open from the dock. Muscle memory already voted.

Related: Integrated Cloud Environment, where we introduced what ICE is. More about what we're building at light-cloud.com.