Cyanogen Co-Founder Unveils Kora, a Linux Desktop OS Built to Run Multiple AI Models

Kora is a standalone system on the Linux kernel and is scheduled to launch in 2027, McMaster told Wired. The startup has partnered with Qualcomm to bring it to Snapdragon PCs next year, and it also supports x86-64 chips, so it should install on many existing Intel and AMD machines.

The pitch is aimed at developers and creators who already move between several AI providers each day. McMaster said the product is for "the AI native, not mass market". Kora is meant to keep context and memory on the user's own machine and share it across tools. Co-founder Franck De Girolami gave the example of a user who designs a project with an OpenAI model, switches to Anthropic's tools to deploy code, and uses a local model to save on token costs. "All your context, all your memory is on your machine," he said.

At the centre of the design is what the company calls a "neural file system", which indexes files by meaning. Users can type questions into a system search bar and find, for example, a contract by a detail they remember rather than by its file name. The company says compatibility layers will let users run standard Windows applications. It is also building its own email, calendar and contacts apps and a software development kit for outside developers.

Kora grew out of Playtron, McMaster's earlier startup, which released a gaming operating system called GameOS last year. That technology now sits inside Kora as GameFrame for running PC games. The third co-founder, John Lagerling, previously worked on Android at Google.

The company has fewer than 100 employees and is entering a market dominated by Microsoft and Apple, and Google launched its Gemini-based Googlebook OS and laptops in October. McMaster argued that the incumbents cannot change quickly. "When you're a billion-user incumbent, it's impossible to tip the boat," he said.

Kora is being optimised for desktops with at least 64 gigabytes of memory, twice the capacity of the top Googlebook, a sign of how much local computing power the company expects users to run their own models on. The company has not announced pricing.