Get started

Quickstart

This guide takes you from bare machines to a live, load-balanced model endpoint. Total time: about two minutes, most of which is the model download.

Prerequisites— one or more machines with a supported GPU (NVIDIA, AMD, or Apple Silicon), on the same network. That's it.

1. Install the agent

Run this on every machine you want in the pool:

$ curl -fsSL clystria.sh | sh

The installer is ~1.2 MB, has zero dependencies, and fingerprints the GPU, VRAM, interconnect and thermals on start. It registers itself as a system service and comes back up after reboots.

2. Forge the cluster

On the first machine, create a named cluster. On every other machine, join it with the token printed below.

$ clystria init home
# ✔ home created · join token: fg_7Kp2…9xQ

$ clystria join --network home --token fg_7Kp2…9xQ

Confirm the pool has assembled:

$ clystria status
▸ home · 3 nodes · 136 GB VRAM · healthy

3. Deploy a model

Point Clystria at any model on Hugging Face or the Clystria registry. It picks quantization, plans layer placement, and exposes an endpoint.

$ clystria deploy llama-3.1-70b --quant awq
→ planning placement across 3 GPUs…
→ sharding · KV-cache balanced · 0 spillover
✔ live at http://home.local:8080/v1

4. Call it

The endpoint is OpenAI-compatible. Point any existing SDK at it — just change the base URL.

$ curl home.local:8080/v1/chat/completions \
  -d '{"model":"llama-3.1-70b","messages":[{"role":"user","content":"hi"}]}'

That's the whole ritual. Add a machine later? Power it on — it fuses into the pool automatically and the scheduler rebalances. Lose a machine? Layers re-forge onto survivors mid-stream.