Private · Open-source · On-premises AI

Your AI. Your data. Your machines.

LAIR is the complete private AI stack you self-host: models, automation and tools running entirely on your own servers. No big player, no lock-in.

✓

Self-hosted

✓

No SaaS lock-in

✓

Fully forkable

I agree to the processing of my data as described in the Privacy Policy

Request a live demo to see LAIR in action. We'll email you a link to schedule your session.

~/projects/lair — zsh

$ git clone https://github.com/Lair-ai/lair.git

Cloning into 'lair'...

remote: Enumerating objects: 3901, done

Receiving objects: 100% (3901/3901), 6.81 MiB done.

Resolving deltas: 100% (2742/2742), done.

$ cd lair

~/lair$ sudo ./setup.sh

The stack

One command. The whole AI

LAIR provisions a full, integrated AI environment on Kubernetes - chat, automation, model serving, document RAG and image generation - all addressable on your own domains.

💬

OpenWebUI

Private ChatGPT-like interface with RAG and extended features

ai.host.local

🔗

N8N

No-code workflow automation & API orchestration, 400+ integrations

automation.host.local

🦙

Ollama

Local LLM serving with GPU acceleration: your models, your hardware

internal API

📄

Tika

Document processing & text extraction for RAG context

internal API

🖼️

ComfyUI

Local AI image generation when a GPU is available

images.host.local

🗄️

MinIO

S3-compatible object storage for files and backups

storage.host.local

🐘

PostgreSQL

Database backend for OpenWebUI and N8N

internal API

⚡

Redis

Cache & queue backend for OpenWebUI and N8N

internal API

CLUSTER & HOST LAYER ingress-nginx cert-manager Let's Encrypt Longhorn Velero CoreDNS MetalLB dnsmasq mkcert

Runs anywhere

From a 60W edge device to a 240W desktop supercomputer.

One production AI stack on ARM64 and x86_64. The installer detects the platform and configures networking, storage and GPU automatically.

PLATFORM	ARCH	NETWORKING	STORAGE	GPU	MULTI-NODE
x86_64	x86_64	Calico	Longhorn	NVIDIA (opt.)	yes
NVIDIA Jetson	ARM64	Flannel	Hostpath	Tegra	single-node
NVIDIA DGX Spark	ARM64	Flannel	Hostpath / Longhorn	GB10 Blackwell	via ConnectX
Cloud instances	x86_64	Calico	Longhorn	NVIDIA (opt.)	yes

DESKTOP SUPERCOMPUTING

Validated on NVIDIA DGX Spark

The GB10 Grace Blackwell Superchip brings datacenter-class AI to a desktop. Run 70B models in FP16 on a single unit; link two via NVIDIA ConnectX to reach 405B+ parameters. DGX OS is Ubuntu-based, so the entire stack runs natively.

128 GB

UNIFIED MEMORY

1 PFLOP

SPARSE FP4

~200B

PARAMS, SINGLE UNIT

273 GB/s

BANDWIDTH

Local-first inference

Zero data egress

Open on release day

Runs anywhere

Why LAIR

Built around data sovereignty, not added on later.

The anti-SaaS AI stack. Everything runs inside your perimeter: no data egress, no vendor in your path.

Privacy-first by default

Prompts, outputs and logs stay on your infrastructure. Nothing is phoned home, verify it in the network tab.

Open-source & forkable

No black boxes. Read every line, fork it, ship it. Released under a permissive license the day it lands on GitHub.

Everything included

Chat UI, workflow automation, local model serving and object storage, all in one stack on your own cluster. No SaaS in your data path.

How to get started

Deploy your private AI stack from GitHub to your terminal.

The anti-SaaS AI stack. Everything runs inside your perimeter — no data egress, no vendor in your path.

CLONE

Clone the repository

Access the public repository on GitHub and clone the project to your local machine or server cluster.

DEPLOY

Run the setup

Execute the quick installation script. It automatically configures the environment and downloads the required local components.

BUILD

Start building

Open the UI, connect your local models, and start automating your workflows with total control over your data.