Startup Kora is building a Linux OS where OpenAI, Anthropic and local models share a common context

A familiar picture: you design a product in an OpenAI model, switch to Anthropic's tools to deploy the code, and run routine tasks on a local model so you don't burn through tokens. Each tool lives in its own tab or desktop app and knows nothing about its neighbors. You carry the context between them by hand. The startup Kora believes this problem should be solved not with yet another plugin, but at the operating system level.

What Kora is

Kora is a standalone OS built on the Linux kernel that runs directly on the hardware. It was announced on Tuesday and is due out in 2027. The company has agreed with Qualcomm to launch on Snapdragon PCs next year. The x86-64 architecture is also supported, so the system can most likely be installed on current machines with Intel and AMD processors.

The key thesis, according to co-founder Frank De Girolami: all context and all memory are stored on your machine. Models from different providers get access to the same data layer, and you choose whom to plug in for a specific task. An expensive cloud model can be used for design, and a local one for high-volume, cheap operations.

To avoid ending up in a vacuum without software, the developers built in compatibility layers for running ordinary Windows applications. In parallel, they are writing their own programs for contacts, calendar and email, and preparing an SDK for third-party developers. At the bottom of the home screen is a familiar dock for switching between email, the browser and games.

The "neural file system"

The second notable idea is what the team calls a neural file system. File data is indexed by meaning, so you can search not by name but by content. An example from the source: you remember one small detail of a business contract but don't know where it is stored or what it's called. You type a description into the OS search bar and get the document you need. Judging by the screenshots, the interaction resembles a conversation with a chatbot or an agent.

The project's roots are in gaming. Kora grew out of the startup Playtron, which last year released GameOS for handheld consoles like the Steam Deck. That architecture now lives inside Kora under the name GameFrame and lets you run PC games if the hardware is powerful enough.

Who is building it and who it competes with

The team has fewer than 100 people, but the founders have launched operating systems before. Kirt McMaster is a co-founder and one of the leaders of Cyanogen, an open-source Android fork that had millions of users until the project shut down about ten years ago. De Girolami took part in developing GameOS, and John Lagerling previously worked on Android at Google.

There is already a competitor: in October, Google released Googlebook OS and compatible laptops tailored for Gemini. McMaster believes that it and its Android support are not suited to power users who need heavy models. Kora is being optimized for configurations with at least 64 GB of RAM, twice as much as the top Googlebook. McMaster's bet: a giant with a billion users cannot change course quickly, while a small team can develop noticeably faster than macOS and Windows.

What to make of it and where the catch is

For now this is an announcement, not a product. There is no release, no independent tests, and no details on exactly how shared context will travel safely between providers. It is unclear how well the compatibility layers will handle real Windows applications. The 64 GB memory requirement immediately rules out most ordinary laptops. The history of alternative operating systems is not especially optimistic either: the main problem is always the app ecosystem.

The practical takeaway for developers: watch the SDK. If Kora takes off, there will be one more platform where agents and applications have access to a single local context. Product managers and founders should note the trend itself: the "memory" and context of AI tools are moving from separate services to the OS level. It's a reason to think ahead about the portability of user context between models and not to build a product on the assumption that it will stay inside a single browser tab forever.