Veda, a Rust-built hobby operating system, puts an AI agent at the center of the OS

Veda is an operating system with an AI agent built into it. The project lives on GitHub under vahmoh25/Veda. Its makers say the agent is not an assistant bolted onto a desktop, but a part of the system that every application is made to work with.

You talk to it in plain speech. The project's own example is "Hey Veda, switch my wallpaper to the dunes." The agent then does the job through the same operations as the keyboard and mouse, and it asks first before anything that is hard to undo.

The code is written from scratch in Rust. That covers the UEFI bootloader (the first program that starts the computer), a capability-based microkernel (a very small core that only handles isolation, scheduling, memory and messaging), the drivers, system services, window system, toolkit and applications. A few existing libraries handle networking, TLS and cryptography: smoltcp, rustls and RustCrypto. Veda runs on x86-64 PCs and in QEMU, a program that emulates a computer.

The project's argument is that operating systems were designed for a person at a keyboard. An assistant added later usually has to read the screen, imitate clicks, or rely on the few apps that offer an API. The system itself has no say in what it does. Veda treats the agent as a first-class user instead.

In practice, the agent runs as a system service. It starts at boot, restarts if it fails, and sits as a small ring on the taskbar. You wake it by saying its name, clicking the ring or pressing Super+Space.

Apps talk to the agent directly. Each app describes its actions, with typed parameters and a risk level, and reports what it shows as structured state. Actions run through the same code as keyboard and mouse input, so there is no screen scraping or simulated clicking. The built-in apps offer more than 70 actions between them, from writing in a document to starting a race.

Consent is enforced by the system, not by the language model. Every action is routine, sensitive or destructive. Routine ones just happen, while the others wait for the user's approval. According to the project, only the desktop shell can give that approval, and nothing the agent reads in a document or app can approve anything. Sensitive actions can be allowed permanently, but destructive ones are asked about every time.

The agent's memory holds only what the user tells it, such as a name, people in their life and how they like things done, and never its own guesses. It is stored on the user's computer, and Settings can show or delete any of it.

Voice comes first. Conversations run in real time, and you can interrupt the agent. The audio system uses echo cancellation so the agent does not hear itself, and it turns other sound down while the agent talks. While the agent sleeps, a voice detector on the computer listens for speech, so that nothing leaves the machine while nobody speaks.

The speech recognition, language model and voice come from Deepgram's Voice Agent platform. The project says the audio is opted out of Deepgram's model improvement. The model, the voice and even the agent's name can be changed in Settings.

The rest of the system is broad for a hobby project. Drivers and services run outside the kernel, supervised by init, so if the window system crashes it restarts and the desktop returns. The desktop has a compositing window manager, a taskbar, a start menu, a calendar and notifications. There are network and Wi-Fi services, a file system with crash-safe snapshots, and apps including a text editor, photos, music, files, terminal, task manager and settings.

Veda also runs Linux drivers inside a virtual machine, on a hypervisor it wrote itself, to support hardware such as USB, networking, displays and GPUs. It includes an OpenGL ES 3.0 graphics stack, two 3D games called Velocity and Starfall, and a demo called Prism. Users can write C programs in the Terminal or Text Editor and compile them with GCC 16 on Veda itself.