Why a desktop AI company, not another chat app

Single-agent chat is the dominant shape today. ChatGPT, Claude, Gemini, Mistral — they all converged on the same UI: one input box, one assistant, one rolling thread of context. The shape is so good for single-task work that for a while it looked like the whole problem of “AI for work” was just going to be solved by making the single assistant smarter.

The second shape is the agent framework. CrewAI, AutoGen, LangGraph — Python SDKs that let engineers wire up multiple agents in code, with planners and message buses and orchestration. They take coordination seriously, but the user is the engineer, and the surface area is a class diagram.

The third shape is the cloud agent platform. Lindy, Devin, Operator — managed services that run agents on someone else's infrastructure against integrations they preconfigured. They cover the customer-facing edge of work that already lives in SaaS, and they cover it well.

What is missing from those three is the consumer desktop app where multi-agent is not an integration project — where you download an app, hire a team, and DM whoever you need without writing code or wiring connectors. That is the shape we picked.

Multi-agent was hard for a long time. Models had to get good enough to follow long instructions, hand off cleanly, and resist drifting. Tool use had to get reliable enough that an agent could do real work without supervision. The frameworks landed on that frontier first because they had engineering audiences willing to debug their way through the rough edges.

Desktop was unfashionable for AI. The first generation of AI products went to the browser because that was where everyone already was, and because the model APIs were a server-side thing anyway. Building a desktop app meant building Electron-shaped infrastructure or learning Swift or signing for two stores, none of which felt worth it when the alternative was a URL.

The combination — multi-agent on the desktop — was nobody's plan. The framework people did not want to ship a consumer app. The single-agent chat people did not want to take on the orchestration. The cloud agent people did not want to give up the integrations. The shape was sitting in the middle, and we picked it.

Tauri made it tractable. The runtime is a small Rust binary that hosts a web view; the UI is React; the file system, shell, and notification APIs are first-class instead of bolted on. The same code that ships to macOS ships to Windows. Local SQLite for storage. The whole stack feels less like fighting the platform and more like the platform was waiting for someone to do this.

The file system. The single biggest reason to ship a desktop product is that the agents can read and write the files you actually work in, without a sandbox VM in the middle. A web chatbot can read what you upload; a desktop agent can read the folder you point it at. The difference is enormous when the work is iterative.

Local storage. Conversations, per-employee memory, hire decisions, and any sandboxed files persist in local SQLite. No cloud sync, no telemetry of prompt or response content, no provider keys baked into the binary. The privacy story is structural rather than policy-based — the data does not leave the machine because there is no infrastructure to leave it to.

A curated shell. Each employee can run commands in a capability-scoped shell — enough to compile a project, run tests, format a file, or check a git status. Destructive actions pause for your OK. This is the missing primitive in every web-only product: an agent that can actually do the thing on the machine you actually use.

Notifications, scheduling, and background work. The OS already has primitives for all of these. A desktop app can use them. A web app fights to approximate them through service workers and tabs that have to stay open.

Desktop costs you mobile. Replace OS is not on your phone. That is a real cost for the kind of casual reaching-for-AI work that ChatGPT owns on iOS. We do not pretend that does not matter; we picked the trade deliberately because the high-leverage work for the audience we serve happens on a laptop, not a phone.

Desktop costs you the browser. The killer feature of any web product is the URL — you share a link and the recipient is in. Desktop apps need installs. We get around the worst of this with single-binary downloads and a 30-second setup, but the friction is still there.

Desktop costs you updates. We push updates regularly during beta, the app handles the download in the background, restarts are quick. None of it is as transparent as refreshing a tab.

What we got back for those costs: the file system, the shell, real local storage, structural privacy, and the ability to build a product where the agents do the actual work instead of describing it. That is the trade.

Five years from now, AI usage is going to fragment into a small number of stable shapes. Single-agent chat will keep most of the casual, single-task usage because the shape is right for the work. Cloud agent platforms will keep the customer-facing automation surface because the cloud is where that work already lives. Frameworks will keep the engineering audience because that is the audience they were built for.

The shape we are building for is the one where a person sits down at a laptop, opens an app, and says “the team I would build if I had a small company, already here.” That is the shape we keep hearing from beta users — and it is the shape that justifies all the trades above.