Skip to content

Getting Started

Jump to bottom
rzfz.ai release bot edited this page Jul 29, 2026 · 1 revision

Getting started

The rzfz.ai Stack — the integrated open-source stack for local AI infrastructure — is designed to go from a bare Linux host to a working AI platform in a single command. This page is the friendly on-ramp: what you need, the quickstart, and what you get. When you are ready for the step-by-step, head to Community install.

Prerequisites at a glance

You need three things:

  • A Linux host. Ubuntu 26.04 LTS is the supported base. Everything runs in containers, so you do not install service packages on the host itself.
  • Docker + Docker Compose v2. Docker Engine 24.0+ with the Compose v2 plugin (docker compose, with a space — not the legacy docker-compose v1 binary).
  • Enough hardware. Plan on 32 GB RAM or more for a useful deployment — the always-on core alone (database, identity, cache, proxy, mail relay, and the built-in UIs) needs well over 16 GB before you load a single model.

For LLM inference the stack supports three hardware targets, chosen at install time and auto-detected if you don't say:

  • AMD (e.g. Strix Halo, with integrated/unified memory)
  • NVIDIA (CUDA GPUs)
  • CPU-only (no GPU — works, but slower; fine for light or batch use)

See Host requirements below for a little more detail.

The three-line quickstart

On a prepared host with the code checked out:

rzfz init                 # 1. run the interactive setup wizard
                          # 2. answer domain / timezone / password / profiles
                          # 3. open the UIs it prints when it finishes

rzfz is the unified command-line tool for the stack — one entry point, rzfz <command>. rzfz init checks prerequisites, generates all secrets, asks for your domain, timezone, admin password, and which modules to enable, then builds and pulls the container images and starts everything.

Prefer a non-interactive install? Pass a preset that bakes in a whole scenario:

# AMD single box, self-signed TLS, common modules (hardware auto-detected)
rzfz init --package single-box --domain myai.local --password 'SecurePass123!'

The full walkthrough — cloning the code, presets, verifying health, opening the UIs — is in Community install.

What you get

When rzfz init finishes and you run post-install provisioning, you have a running AI platform reachable on subdomains of your base domain, all behind single sign-on:

  • Chat (chat.<domain>) — Open WebUI, talking to local models.
  • Workflows (dify.<domain>) — Dify for building LLM apps.
  • LLM management (llm.<domain>) — GPUStack, where your models load and run.
  • Identity (auth.<domain>) — Authentik, your single sign-on and user management.
  • Configuration portal (config.<domain>) — enable modules and change settings after install.
  • Help (help.<domain>) — in-product documentation, served offline.

Plus the optional modules you enabled — monitoring, private web search, speech, document conversion, agents, and more.

First-run pointers

  • Open the UIs behind SSO. Every app sits behind Authentik. The first time you visit one, you log in through Authentik and it hands you to the app. See Architecture overview for how that gate works.
  • Deploy the default models. A fresh install has the apps but no models yet. Post-install provisioning deploys a sensible default model set (a general chat model, embeddings, and a reranker) so chat and retrieval work out of the box.
  • Enable more modules any time. Modules are Compose profiles; turn them on from the Configuration portal or by editing your profile list.

Host requirements

A little more detail on the "enough hardware" note above:

You are running… Typical RAM GPU / VRAM Free disk
A single AMD box (integrated/unified memory) 128 GB unified is the reference appliance; core reserves a ~32 GB container budget, the rest is model memory integrated (unified) 250 GB+ NVMe
A CPU-only host or VM (no GPU) 32 GB+ for the core, plus model RAM on top (CPU inference runs in RAM) none 100 GB+
An NVIDIA / CUDA host 32 GB+ host RAM for the containers VRAM by your GPU budget 100 GB+

Notes:

  • Ubuntu 26.04 LTS is the supported operating system. The stack is tightly bound to its Docker / Docker Compose architecture and assumes this base.
  • All apps are reached through Caddy on ports 80/443. Database and internal service ports bind to 127.0.0.1 only. For public DNS + automatic TLS you'll point A records (or a wildcard *.<domain>) at the host; for a purely local install, /etc/hosts entries are enough (self-signed TLS).

The complete per-hardware sizing matrix and the full support statement live in the Enterprise documentation at docs.rzfz.ai.

Next

Clone this wiki locally