Build an on-prem AI workstation — the stack, and the order it goes on.
Everything a machine needs to run local models alongside cloud AI agents — installed in the one order that actually works. Six layers, a verification step at every boundary, and Workspace Guard as the neutral coordination layer on top.
The architecture, top to bottom
You work through one pane of glass. Underneath, two worlds — cloud agents and your local model server — never talk to each other directly. The coordination layer is what keeps them from colliding in the same workspace.
- Layer 5 · You. Your repositories and project files — on the data drive, never the system drive.
- Layer 4 · Coordination (neutral) — Workspace Guard (in development). A single binary with a localhost dashboard: session locks, an agent activity registry, a prompt library, and auto-update. It treats Claude, Codex, and Ollama identically.
- Layer 3A · Cloud engines. The CLIs for Claude Code and OpenAI Codex. The intelligence stays with the provider; only the CLI lives here, each with its own login.
- Layer 3B · Local engine. Ollama and your model library on 127.0.0.1 — working even with the internet unplugged.
- Layer 2 · Runtimes. Git, Node.js LTS, a real terminal (and WSL2 + Ubuntu on Windows). The plumbing everything above assumes.
- Layer 1 · Machine & OS. A clean, fully-updated OS, the GPU driver, and split storage. The readiness scan says ready or not-ready before anything installs.
Layer 4 is Workspace Guard, still in development. Everything below it is proven on real hardware.
Install in order — some doors only open once
Three of these steps are one-way: the storage layout, moving the WSL disk, and pulling models. Do them out of order and you redo gigabytes of work — so the order isn’t a suggestion.
- 0 · Preflight — the machine proves itself. OS fully updated, GPU driver in, data drive created. Run the free AI Readiness Scan: it reports ready / not-ready and why, and never changes a thing.
- 1 · Foundations (~15 min). Git, Node.js LTS, a terminal. On Windows, enable WSL2 + Ubuntu and move its disk to the data drive while it’s still small; redirect your user folders there too.
- 2A · Cloud engines — Claude & Codex (~10 min, manual by design). Install the two CLIs, then log in through the browser. Their accounts, their credentials — never ours.
- 2B · Local engine — Ollama (~10 min, proven recipe). Install Ollama, set it to auto-start, and confirm the API answers on 127.0.0.1:11434 after a full restart.
- 3 · Models (~30–60 min, verified on 12 GB VRAM). Pull the tier that fits your GPU, then verify: every chat model runs 100% on the GPU and answers a real prompt. A model that quietly spills to the CPU is a problem waiting to happen.
- 4 · Workspace Guard (in development). The coordination layer, built to install in five minutes: one signed binary, no admin rights, delete-the-folder uninstall. Point its dashboard at your workspace roots and it discovers the engines above. This is the piece we’re still building.
- 5 · Ongoing — the machine stays ready. A weekly readiness scan that reports drift and never auto-deletes; components self-update. Your backups stay yours, and we say so plainly.
The three rules that can’t bend: storage layout before any install; move the WSL disk before pulling models; verify at every boundary before calling a phase done.
Models sized to your GPU
The 12 GB build is our proven reference — an RTX 3060 with 32 GB RAM, every model verified running fully on the GPU. Neighboring tiers follow the same shape and are awaiting their own verification pass.
| Role | 8 GB | 12 GB (proven) | 16–24 GB |
|---|---|---|---|
| Fast / small | llama3.2:3b | llama3.2:3b | llama3.2:3b |
| General chat | qwen3:8b | qwen3:8b | qwen3:8b |
| Max quality | — (8b ceiling) | qwen3:14b | qwen3:32b class |
| Coding | qwen2.5-coder:7b | qwen2.5-coder:7b | qwen2.5-coder:14b+ |
| Embeddings | nomic-embed-text | nomic-embed-text | nomic-embed-text |
| Disk footprint | ~12 GB | ~21 GB | ~45+ GB |
Verify at every boundary
“Done” means proven, not assumed. The end-state checks:
- Ollama survives a restart — after a full shutdown, the API answers on :11434 on its own.
- GPU residency — the model reports 100% GPU, not a CPU/GPU split.
- Real inference — a model answers a real prompt, from the GPU, in seconds.
- Cloud agents — Claude and Codex print their versions and stay logged in.
- Storage discipline — everything heavy lives on the data drive.
- Coordination — the Workspace Guard dashboard shows all engines with locks working (once it ships).
Start where the stack starts
Phase 0 is a free scan that tells you whether your machine is ready — read-only, nothing uploaded, nothing changed. Everything else builds on it.
Future Maven — Bridging the divide from scarcity to abundance.
