Veda: an agent-native operating system written from scratch in Rust

An operating system built around an agent

Veda is a new open-source operating system, written from scratch in Rust, that treats an AI agent as a core system component rather than an application bolted on afterwards. According to the project's GitHub repository, which surfaced on Hacker News's front page, the agent is a system service started at boot and supervised so that it restarts if it fails. Users wake it by saying its name, clicking a small ring on the taskbar, or pressing Super+Space, after which it acts across applications through the same internal operations the keyboard and mouse use, pausing to ask permission before anything that is hard to undo.

Agent support as a platform primitive

The repository's central argument is that existing operating systems were designed for someone sitting at a keyboard, so assistants layered on later end up reading pixels, faking input, or depending on the handful of applications that expose APIs. Veda inverts that arrangement. Agent capability is part of the application platform itself: each program declares its actions, each with typed parameters and a risk classification, publishes what it displays as structured state, and executes agent requests through the same code path as direct input, so the window shows the outcome exactly as though the user had done it manually. The bundled applications expose more than 70 actions between them, ranging from writing in a document to starting a race in a game.

Consent enforced below the model

Every action falls into one of three classes: routine, sensitive or destructive. Routine actions run immediately; the others are held until the user approves. Crucially, enforcement lives in the agent service rather than the language model. A pending call waits for an answer, only the desktop shell can grant one, the system recognises every program by the identity init attaches to its connections, and nothing the agent reads inside a document or application can authorise anything. Sensitive actions can be allowed permanently, while destructive ones are re-confirmed every single time.

Voice, context and memory

Veda is voice-first. Conversations run in real time, the agent stops and listens when interrupted, and the audio service provides echo cancellation that keeps the agent from hearing its own voice, plus ducking of other sound while it speaks. While dormant, an on-device voice detector listens for speech, so nothing leaves the computer until someone talks. Speech recognition, the language model and the voice come from Deepgram's Voice Agent platform, with the audio opted out of Deepgram's model improvement; the model, the voice and even the agent's name are all selectable in Settings. The agent's memory stores only what the user plainly tells it, never its own inferences, keeps that data locally, and Settings displays and can delete every entry. Each conversation begins with real context: the time, the windows on screen, every application's available actions, and how the previous session ended.

The Rust stack underneath

Everything else is original code as well: a UEFI bootloader, a capability-based microkernel called vkernel that confines itself to isolation, scheduling, memory management and IPC, plus drivers, system services, a compositing window manager, a custom vui toolkit and the applications. If the window system crashes, init restarts it and the desktop recovers without user intervention. The desktop composites frames on the GPU with OpenGL ES on machines with Intel graphics and on the processor elsewhere. Networking is a user-space service covering IPv4 and IPv6, DHCP, DNS, routing, TCP and UDP, with Wi-Fi supporting WPA2, WPA3 and open networks; Wi-Fi relies on Linux drivers running inside a virtual machine and restricted to moving frames. Only a few mature crates are borrowed — smoltcp for TCP/IP, rustls for TLS and RustCrypto for cryptography. Veda targets x86-64 PCs and QEMU, the latter offering a simulated Wi-Fi environment bridged to the Internet.

Why it matters

The industry's dominant approach layers agents onto Windows, macOS and Linux and accepts the friction of screen-reading and synthetic clicks. Veda is a clean-room experiment in the opposite direction: an operating system whose application model presumes an agent from the start, with typed actions, structured state and consent enforced beneath the model rather than trusted to it. It remains a hobby project that leans on a driver VM and cloud-hosted intelligence, and it has yet to prove everyday reliability. But as an inspectable reference design for agent-first computing, it sketches what the desktop might become if agent interfaces were treated as a platform primitive, the way the graphical interface once was.