The first native Ring-0 OS for the AI era

AI runs
at Ring-0.
No host OS.
No syscall overhead.

AgentOS is a purpose-built operating system where the AI is the kernel — not an application running on top of Linux. Executing LLMs directly on bare metal with deterministic safety and machine-speed execution.

149.3
tok/s
CPU-only inference on AMD EPYC
10x
faster
than comparable llama.cpp runs
0
host OS
bare metal, Ring-0 execution
AI Kernel
Ring-3: User Apps
Ring-1: Guest OS
Ring-0: AI Kernel
The Problem

For 50 years, operating systems were built for humans typing into terminals.

The architectures we inherited — kernel/user space splits, syscall overhead, process boundaries — were designed for a different computing era. They are structurally incompatible with how modern AI agents need to operate.

Running AI on Linux is like running a race car on a dirt road. The car is fast. The road is the bottleneck. AgentOS is the road built for the car.

Legacy Stack
Application Layer
Host OS (Linux/Windows)
Virtualization / Hypervisor
Hardware
4 context switches per inference call
AgentOS
AI Kernel (Ring-0)
Hardware Abstraction Layer
Bare Metal
Zero syscall overhead
Architecture

No user-space. No kernel-space split.
No process boundaries. No syscalls.

RING-3 User Space Sandboxed Agents AgentLang Runtime App Binaries (AOT) Type-safe execution via Validator + Capability system RING-0 AI Kernel — The OS Core Boot-LLM Qwen3-1.7B Q4_K_M 149.3 tok/s (CPU-only) x86_64 + ARM64 deterministic Validator Type-safe code execution 270+ test cases passing No runtime JIT overhead Arena Allocators O(1) bump-pointer Zero-Copy IPC Capability-typed handles AOT Codegen Native x86_64 Hardware Abstraction Layer — x86_64 · ARM64 HVF
Safety enforced by Validator and the type system — not page tables
Single Ring-0 address space — no kernel/user split
Unikernel architecture — one address space, one purpose
Licensing

Open core. Two editions.

AgentOS is free for individuals and hobbyists. AgentOS Server is the paid edition for businesses and production deployments. No hidden costs, no usage limits on the free tier.

Feature
AgentOS
Free
For individuals & hobbyists
AgentOS Server
Paid
For businesses & production
License
AGPLv3
Commercial
Use case
Personal / hobby / research
Production workloads
Supported hardware
Consumer PCs, laptops
AMD EPYC, Intel Xeon, native NICs
SMP / multi-core
Limited
Full SMP enabled
Inference performance
Standard
Up to 149.3 tok/s on EPYC
Support
Community (GitHub)
Priority email + SLA
Compliance builds
—
Certified for regulated industries
Source code
Public (AGPLv3)
Available
Pricing
$0 — always free
Contact for pricing
AgentOS Basic + AgentOS Hardcore are 100% free, forever. The free tier is not a trial — it ships complete. AgentOS Server adds enterprise support, compliance builds, and production-grade tooling.
Benchmark

Real hardware. Real numbers.
Not a slide deck.

AgentOS Server on AMD EPYC 9354P
64 logical CPUs · Native NIC · SMP enabled · No GPU
149.3
tok/s
Qwen3-1.7B Q4_K_M — 32-token validation run, CPU-only
System / Runtime
Model
Mode
Result
AgentOS Server — AMD EPYC 9354P
Qwen3-1.7B Q4_K_M
CPU-only, Ring-0
149.3 tok/s
Fast desktop CPU, llama.cpp / Ollama class
Qwen3-1.7B Q4_K_M
CPU-only
10–15 tok/s
Intel i7-13700K, llama.cpp Q4
Qwen3-1.7B
CPU-only
35 tok/s
Consumer GPU, 6 GB VRAM
Qwen3-1.7B
GPU
~126 tok/s
Qwen official SGLang benchmark, 1 GPU
Qwen3-1.7B BF16
GPU serving
227.80 tok/s
GPU acceleration is the next target. The architecture doesn't exclude it — it delays it until the Ring-0 foundation is solid. Stage 16+.
"
We don't sandbox the AI inside a guest.
The AI is the kernel.
Stage 11
Boot-LLM Operational
Cross-platform deterministic. Validated on AMD EPYC 9354P. CPU-only inference at 149.3 tok/s.
Stage 12
Sandboxing + AgentLang Migration
Full sandbox isolation between agents. AgentLang migration from experimental to production runtime.
Stage 14
Dual Desktop — Human Oversight
Classical user desktop alongside Agent Desktop: live visualization of agent reasoning, state, and intent — in human-readable form. Pillar 2: Human Oversight, operationalized at the OS level.
Stage 24+
Pure-AgentLang Kernel
The AI kernel written in AgentLang itself. Self-hosting. The OS that runs itself.
Open Source

The kernel is free. It stays free.

AgentOS Basic for the everyday developer, AgentOS Hardcore for builders and hackers, and the open AgentOS Server foundation are AGPLv3. The kernel source code will be published to this repository under AGPLv3. Early access is available to select design partners and core contributors.

View on GitHub
Product Family
AgentOS Basic
Free
Everyday local AI OS, AGPLv3
AgentOS Hardcore
Free
Builder and hacker edition
AgentOS Server
Open Foundation
EPYC/Xeon, native NICs, CPU/GPU
Certified Editions
Enterprise
Compliance builds for regulated industries
AgentOSLLM Cloud
Sovereign
Cloud reasoning when local hardware isn't enough

The next operating system
has already started.

AgentOS is not a roadmap. It is shipped code running on real server hardware. The kernel is open. The vision is clear. The next era of computing has a new OS.

Stage 11 of 24+ — and accelerating.