Veda Introduces an Agent-Native Operating System
Summary
Veda is a Rust-based operating system designed with a voice agent as a first-class system service rather than an assistant layered on top of a conventional desktop. The agent starts with the system, can be awakened by name, a taskbar control, or Super+Space, and uses structured application actions instead of screen scraping or simulated clicks. Applications expose typed actions and JSON state, while the built-in applications provide more than 70 actions across windows, files, media, networking, settings, and other functions. Veda classifies actions as routine, sensitive, or destructive; its agent service, rather than the language model, controls approval for sensitive and destructive operations, with destructive actions requiring confirmation each time and approvals expiring after three minutes. The system supplies conversation context from the time, open windows, available actions, user-controlled memory, and the previous conversation, while memory remains on the computer and can be reviewed or deleted in Settings. Voice detection runs locally while the agent sleeps, and Veda uses Deepgram's Voice Agent API for recognition, turn-taking, language-model interaction, and speech; OpenAI's gpt-4.1-mini is the default model, but the model, voice, and agent name are configurable. Deepgram audio is opted out of model improvement, according to the README. The project includes a capability-based Rust microkernel, user-space services, a Linux driver VM, desktop applications, networking, audio, video, and OpenGL ES 3.0 support. It runs on x86-64 PCs and in QEMU, with a live USB image also available, although real-device support depends on hardware and virtualization features. The repository provides deterministic offline agent tests using a stand-in for Deepgram, alongside system, UI, networking, crash-recovery, and application-action tests. Developers can make applications agent-compatible by declaring actions, exposing state, and routing agent calls through the same implementation used by keyboard and mouse input.