⚠️ WARNING: These scripts are functional but far from ready for general use. You may encounter rough edges, incomplete error handling, or sharp corners. Proceed at your own risk.
⚠️ AMD RX 7000 Series and Newer Users: Driver 26.3.1 is recommended for stability across all workloads — ROCm, desktop productivity, and gaming. Some users report instability with 26.5.x and newer. See KNOWN_ISSUES.md for details.
Provision a structured local AI workstation on Windows in three scripts. Models, runtimes, config, workspace, tools, and caches each get their own layer — no duplication, safe reinstalls, no lost models.
D:\AI\
├── AI_CONFIG\ settings, model registry, port config
├── AI_CORE\ runtimes (ComfyUI, Ollama, Open Web UI)
├── AI_VAULT\ permanent models (single source of truth)
├── AI_WORKSPACE\ input, output, workflows
├── AI_TOOLS\ helper scripts, launchers, utilities
└── AI_CACHE\ temp data (safe to delete anytime)
Core principle: models live once in the vault. Every runtime consumes them through a symlink binding layer. Reinstall any tool without touching your data.
- Windows 10 or 11
- PowerShell 5+
- Administrator rights (for
2-deps.ps1)
-
Download and unzip the latest release — you'll get a
scripts/folder with the four bootstrap scripts. -
Open PowerShell in the unzipped folder and run:
# Allow local scripts
Set-ExecutionPolicy -Scope CurrentUser RemoteSigned
# Unblock downloaded files (if downloaded from the web)
Get-ChildItem *.ps1 | Unblock-File
# Step 1 — create folders, symlinks, config
.\scripts\1-init.ps1
# Restart PowerShell
# Step 2 — install Git, Python, Ollama, FFmpeg
.\scripts\2-deps.ps1
# Restart PowerShell
# Step 3 — install apps (ComfyUI + Open Web UI)
.\scripts\3-apps.ps1
# Make the ai command available everywhere
.\scripts\ai.ps1 setup path| Command | What it does |
|---|---|
ai start all |
Start all services |
ai stop all |
Stop all services |
ai restart all |
Restart all services |
ai status |
Dashboard — running services, ports, model counts |
ai doctor |
Full system diagnostics |
ai list |
List installed models by category |
ai install all |
Install or update everything |
ai install <service> |
Install or update comfyui, comfyui-manager, ollama, openwebui |
ai watch <service> |
Live-tail service logs (comfyui, ollama, openwebui) |
ai setup ports |
Change service ports |
ai setup env |
Check and fix environment variables |
ai clean cache |
Free up disk space — pin models to AI_VAULT first (HF, PyTorch, ComfyUI temp, logs) |
Tip: Due to the nature of ComfyUI's frequent release cycle, running
ai install comfyui(orai install all) periodically pulls the latest version and updates dependencies. Most issues resolve with an update.If a service doesn't appear to start, run its launcher directly in a visible PowerShell window to see real-time output:
D:\AI\AI_TOOLS\launch_comfyui.ps1Or use
ai watch <service>in a second window while starting.
| Script | Purpose |
|---|---|
scripts/1-init.ps1 |
Creates folder structure, detects GPU, sets up symlinks, writes config files |
scripts/2-deps.ps1 |
Installs Git, Python 3.10/3.11, Ollama, FFmpeg via winget. Sets environment variables |
scripts/3-apps.ps1 |
Installs ComfyUI + Open Web UI (clone, venv, model paths, launchers). GPU-aware backend selection |
scripts/ai.ps1 |
Daily driver CLI — service management, status, diagnostics, model listing, cache cleanup, environment setup |
The scripts detect your GPU and install the correct backend automatically:
- NVIDIA — CUDA (via standard torch)
- AMD — DirectML or ROCm. DirectML is the default compatible backend; ROCm
(native PyTorch, Python 3.12) is available for faster inference on AMD GPUs
with driver 26.2.2+. Select during install or pass
--switch rocm. ROCm supports RDNA2+ (RX 6000/7000/9000 series); RX 5000 series uses DirectML. See COMPATIBILITY.md for details. - Intel CPU / no GPU — CPU-only fallback (PyTorch without CUDA or DirectML). Slow but functional.
| Service | Default Port |
|---|---|
| Ollama | 11434 |
| ComfyUI | 8188 |
| Open Web UI | 3000 |
Change with ai setup ports. Settings persist in AI_CONFIG\ports.json.
The scripts redirect download caches away from the vault:
| Variable | Points To | Purpose |
|---|---|---|
OLLAMA_MODELS |
AI_VAULT\models\llm |
Ollama model storage |
HF_HOME |
AI_CACHE\huggingface |
Hugging Face downloads |
TORCH_HOME |
AI_CACHE\torch |
PyTorch cache |
Check and fix with ai setup env.
The 6-layer structure separates configuration, runtimes, assets, workspace, tools, and caches into independent layers:
AI_CONFIG\ — centralized configuration and model registry
AI_CORE\ — AI runtimes and applications
AI_VAULT\ — permanent models (LLMs, diffusion, embeddings, CLIP)
AI_WORKSPACE\ — user files (input, output, workflows)
AI_TOOLS\ — helper scripts and utilities
AI_CACHE\ — temporary downloads, caches, and logs (safe to delete)
A symlink binding layer at AI_CORE\_bindings routes model requests from
runtimes to AI_VAULT\models\, keeping the vault as the single source of truth.
- Bootstrap scripts guide
- AI Control Panel reference
- Folder architecture rationale
- FAQ / troubleshooting
MIT