How to Run a Full Local AI Stack on Your Mac — Free, Private, No Subscriptions

Before You Start — Do These First

These accounts take a few minutes but will block you mid-setup if you skip them:

1. Hugging Face account (free) — go to huggingface.co, sign up, then go to FLUX.1-schnell → accept terms. Then Settings → Access Tokens → New Token → copy it. You'll need this later.
2. Docker Hub account (optional) — hub.docker.com. Only needed for future use.

Hardware Requirements

The Three Tools — What Each One Does

Before diving into setup, here's how the three components relate to each other:

LM Studio — The engine. It downloads AI models (Qwen, Llama, Mistral) into your Mac's RAM and runs them locally. Instead of sending your prompts to ChatGPT, you send them to your own model on your own hardware. It also runs as an API server on port 1234. Think of it as the car's engine — nothing moves without it.

Open WebUI — The steering wheel. A ChatGPT-like browser interface that connects to LM Studio in the background. Gives you chat history, system prompts, separate accounts, knowledge bases — everything that makes it feel like ChatGPT but entirely local. Without LM Studio it does nothing. Without Open WebUI, LM Studio has no face.

ComfyUI — The image studio. Completely separate from the chat stack. Loads Stable Diffusion / FLUX models and generates images and video through a node-based interface. Run it independently — it does a completely different job.

Mac (Apple Silicon)
│
├── LM Studio          ← runs AI models (LLM)
│     └── API port 1234
│
├── Docker Desktop
│     └── Open WebUI   ← ChatGPT-style interface (port 3000)
│           ├── account: admin (you)
│           ├── account: user1
│           └── account: user2, user3
│
└── ComfyUI            ← images / video (port 8188)

Step 1 — Install Docker Desktop

Docker runs Open WebUI in a contained environment. Think of it as a box that keeps everything organized. You don't need to understand it — just install it like any Mac app.

1. Download from docker.com/products/docker-desktop — Apple Silicon version.
2. Install normally. Accept all permission requests.

After installing, two critical settings:

• Settings → General: Enable "Start Docker Desktop when you sign in." Without this, Open WebUI won't load after a reboot.
• Settings → Resources: Disable "Resource Saver." If left on, Docker pauses after inactivity and appears broken.

Step 2 — Install LM Studio and Download Models

Download from lmstudio.ai and install. Open it — you'll see a Discover tab (like an app store for AI models).

If you have 64 GB+ RAM

If you have 16–32 GB RAM

Search each model name in the Discover tab, hit Download, and wait (they're large — download overnight on slow connections).

Once downloaded: go to the Developer tab → Start Server. You'll see Server running on port 1234. Then enable "Serve on Local Network" so other devices on your WiFi can access it.

Step 3 — Install Open WebUI

Open Terminal (find it with Spotlight — Cmd+Space → "Terminal") and run this single command:

docker run -d -p 3000:8080 \
  --add-host=host.docker.internal:host-gateway \
  -e OLLAMA_BASE_URL="" \
  -e OPENAI_API_BASE_URL="http://host.docker.internal:1234/v1" \
  -e OPENAI_API_KEY="lm-studio" \
  -v open-webui:/app/backend/data \
  --name open-webui \
  --restart always \
  ghcr.io/open-webui/open-webui:main

What this does: tells Docker to download Open WebUI, run it on port 3000, connect it to LM Studio on port 1234, and auto-restart if anything crashes.

Wait 2–3 minutes. Then open your browser and go to http://localhost:3000. You'll see a ChatGPT-style interface. Create your first account — the first user automatically becomes Administrator.

Step 4 — Set Up Accounts

Each person gets their own account with separate chat history. Go to Settings (gear icon) → Admin Panel → Users → + Add User:

Step 5 — Create Model Profiles

Instead of exposing model names like qwen3.5-35b-a3b-uncensored to users, create friendly named profiles. Go to Workspace → Models → + New Model:

• 🤖 Assistant — Base: Qwen 3.5 35B. System prompt: "You are a helpful assistant. Answer clearly, concisely, and accurately."
• 💻 Code — Base: Devstral Small 24B. System prompt: "You are an expert developer. Write clean code and explain your reasoning."
• ⚡ Quick — Base: Meltemi 7B. System prompt: "Answer briefly and directly. No unnecessary explanation."

Set 🤖 Assistant as the default in Admin Panel → Settings → Default Model. Users open the app, type, and get an answer — no model-picking required.

Step 6 — iPhone and iPad Access

Every device on your network can use the AI. First, find your Mac's local IP:

ipconfig getifaddr en0

You'll get something like 192.168.1.100. On any iPhone or iPad on the same WiFi, open Safari and go to http://192.168.1.100:3000.

Make it a proper app: tap the Share button → "Add to Home Screen." It opens fullscreen, exactly like a native app. Nobody will know it's running locally on your Mac.

Step 7 — ComfyUI for Image Generation

Skip this step if you only need chat. ComfyUI is a separate tool that handles AI image and video generation — completely independent from the chat stack.

Open Terminal and run these commands in order:

brew install python@3.11

(If brew isn't installed, go to brew.sh first.)

cd ~ && git clone https://github.com/comfyanonymous/ComfyUI.git

cd ~/ComfyUI && pip3.11 install -r requirements.txt

cd ~/ComfyUI && python3.11 main.py --force-fp16

Look for Device: mps in the output — this confirms it's using your Apple Silicon GPU. Then open http://localhost:8188.

Step 8 — Download FLUX.1 (Image Model)

Login to Hugging Face (use the token from Step 0):

python3.11 -c "from huggingface_hub import login; login('YOUR_TOKEN_HERE')"

Download the FLUX.1 quantized model (~7 GB instead of 24 GB — same quality output):

cd ~/ComfyUI/models/checkpoints
python3.11 -c "
from huggingface_hub import hf_hub_download
hf_hub_download(
    repo_id='city96/FLUX.1-schnell-gguf',
    filename='flux1-schnell-Q4_K_S.gguf',
    local_dir='.'
)
print('Done!')
"

Then in the ComfyUI interface, open the "1.1 Starter – Text to Image" template → click "See Errors" → "Download all." This auto-downloads ~8 GB of helper files (text encoders + VAE).

What You're Saving

The hardware (Mac) pays for itself in under two years in saved subscriptions alone — before accounting for privacy, speed, and no rate limits.

What's Next (Phase 2)

• Video generation: Wan2.1 or CogVideoX-5B via ComfyUI
• Music generation: MusicGen Large (Meta) via ComfyUI
• Browser automation: OpenClaw + local LLM for automated tasks
• Workflow automation: n8n for morning digests, alerts, and automations
• RAG on your own notes: Ask the AI about your own documents via Open WebUI RAG