From scratch. In Rust. For AI.

What if your
OS was built
for agents?

Oxide OS is a microkernel where AI agents aren't guests running on top — they're the reason the kernel exists.

scroll to explore ↓
Chapter 01 — The problem

Every agent today runs on
an OS designed for humans

You deploy 50 AI agents. They need isolation, communication, supervision. Linux gives them none of it natively. So you bolt on Docker, Redis, Kubernetes, iptables — layers on layers of workarounds.

No real isolation

A compromised agent in Docker can still probe the network, read environment variables, access shared volumes. Containers are walls with holes.

No native IPC

Agents need three external services (Redis, RabbitMQ, gRPC) just to send each other messages. For an OS primitive.

No supervision

Agent crashes. Nothing restarts it with the right context. Kubernetes can restart pods, but it doesn't understand agent hierarchies.

Security teams block deployment

"How do we let AI act autonomously without giving it the keys to the kingdom?" Nobody has a good answer.

Chapter 02 — The answer

Make the kernel
understand agents

Don't add agent support on top. Build it in. Every kernel primitive — scheduling, memory, security, communication — designed for autonomous AI.

01

Capability-based security

Every resource access requires an unforgeable kernel token. An agent with zero capabilities can do nothing. Delegation enforces subsets only. Revocation cascades instantly.

02

Built-in IPC

Message passing, shared memory, pub/sub channels, request/reply with timeouts. All kernel-native. All capability-gated. Zero external services.

03

Supervision trees

Erlang-style crash resilience. RestartOne, RestartAll, Escalate policies. Agents form hierarchies. Parents control children's fate.

04

Real hardware drivers

PCI bus discovery, virtio-net with split virtqueues, virtio-blk with DMA ring buffers. Real kernel, real I/O, real packets.

05

Pure Rust. No C.

Memory safety at every layer. unsafe only at hardware boundaries. No garbage collector. No runtime. 5,384 lines.

06

Agent-native kernel

AgentConfig with system_prompt, model binding, tools, capabilities, restart policy. Agents are kernel primitives, not userspace processes.

Chapter 03 — The proof

Don't take our word for it.
Watch it boot.

Every line below is real output from QEMU. Hit play.

oxide-os — serial console
Waiting for boot...
Chapter 04 — Under the hood

Peel back each layer

Click any layer to explore. From agents down to bare metal.

AGENTS Agent Swarm +
Supervisor
Spawn, monitor, restart children
Supervision Tree
RestartOne / RestartAll / Escalate
Agent Config
system_prompt, model, tools, caps
Lifecycle
spawn → run → kill → cleanup
KERNEL Kernel Services +
Scheduler
3-level priority, preemptive, fair
Capabilities
Unforgeable tokens, delegate, revoke
IPC
Messages, shared mem, pub/sub, req/rep
Network
smoltcp TCP/IP, HTTP, DNS, firewall
Storage
OxideFS, block cache, KV store
Crypto
HMAC-SHA256, RNG, agent signing
CORE Kernel Core +
Memory
Bitmap frames, 4-level paging, heap
Syscalls
20 defined, x86_64 MSRs, asm entry
Timer
APIC ~400Hz, TSC clock, deadlines
GDT / IDT
Ring-0 + Ring-3, exception handlers
HARDWARE Bare Metal +
PCI Bus
Full enumeration, BAR, bus mastering
virtio-net
256-entry split virtqueues, DMA
virtio-blk
3-desc chains, sync block I/O
APIC / UART
MMIO timer, 16550 serial output
Chapter 05 — By the numbers
5,384
lines of Rust
<1s
boot time
15
kernel modules
20
syscalls
0
lines of C

The infrastructure
for autonomous AI

Open source. Real kernel. Built for what comes next.

View on GitHub Get in touch