Introduction

Clystria is an operating system for AI clusters. You install a lightweight agent on any machine with a GPU — a gaming rig, a workstation, a rack of servers — and they self-assemble into a single, unified AI computer. One pool of VRAM. One pool of compute. One command to deploy a model across all of it.

There is no control plane to babysit, no Kubernetes to learn, and no YAML museum to maintain. If two machines can see each other on the network, they can be forged into one.

The mental model

Think of your machines the way an OS thinks of CPU cores: individually uninteresting, collectively a computer. Clystria abstracts a fleet of heterogeneous GPUs into four primitives:

• Nodes — any machine running the agent. It contributes its GPUs, VRAM, and bandwidth to the pool.
• The Pool— the sum of all node resources, addressable as one. A 405B model doesn't care that its layers live on six different computers.
• Deployments — a model you asked Clystria to run. The scheduler decides which layers live on which GPU.
• The Control Loop— the daemon that continuously rebalances, heals, and routes. It is the "OS kernel" of your cluster.

Why an OS, and not Kubernetes?

Kubernetes was built to schedule stateless web containers. GPUs, model weights, KV-caches and tensor-parallel groups are none of those things. Bending k8s to serve a sharded LLM means operators, device plugins, custom schedulers and a permanent on-call rotation.

Clystria is purpose-built for one job — running models on the metal you own — and it optimizes for that job end to end. The result is a system a single person can operate, that behaves like a cloud you happen to own.

Where to go next

• Quickstart — from zero to a live model in under two minutes.
• Core concepts — nodes, pools, placement, and self-healing in depth.
• CLI reference — every forge command and flag.