Set up your studio
From installation to a first design question. Choose where the model runs: on your own graphics card, or through a cloud route you connect. Atelier itself ships no model.
Download the public source from GitHub, then follow these setup steps. The quick-start guide in the source covers every setting.
Install Atelier
You need Node.js 24 and pnpm 10 or newer, the Atelier folder, and a model to talk to (part two).
- Install Node.js 24, then pnpm:
npm install -g pnpm. - In the Atelier folder, install the packages. This fetches JavaScript only; no model is downloaded.
pnpm install --frozen-lockfileConnect a model
On your graphics card
Start any OpenAI-compatible server (for example llama.cpp, Ollama, LM Studio or vLLM) and note two things: its address, ending in /v1, and the model id it lists at /v1/models. A vision-capable model lets Vitruvius read the images you attach.
No suitable card
Run Atelier with no local model: only Atelier and your design software run on your computer. Connect free tiers or your own API keys through OmniRoute, then start with ATELIER_PROFILE=cloud. Your chats then go to the providers you connect, so keep confidential client work on a local model.
A modest computer
On a small model (about 14B parameters or fewer, or a 32K context) Atelier starts you in Vitruvius Zero, the lightest profile. You can switch profiles at any time.
Which model for which card
| Graphics memory | What to run | How we know |
|---|---|---|
| 32 GB | A 27B vision model with long context. This is what we build and test Atelier on: one Intel Arc Pro B70. | Measured |
| 16 GB | The same family of model in a smaller build, with a shorter context. Slower than 32 GB. | Measured |
| 8 to 12 GB | A 7 to 9B vision-capable model at 4-bit. Start with a 32K context. | Guidance only |
| No graphics card | The cloud route above. | Your providers’ limits apply |
Starts on as little as one 16 GB graphics card: on one card the models take turns, each loading when its job comes up; more cards or a model server keep them loaded side by side.
Start and ask
Give Atelier your model’s address and id on the first run. It writes them into your home folder, so later runs need only pnpm start.
ATELIER_BASE_URL=http://localhost:8000/v1 ATELIER_MODEL=your-model-id pnpm startWindows PowerShell:
$env:ATELIER_BASE_URL = "http://localhost:8000/v1"
$env:ATELIER_MODEL = "your-model-id"
pnpm start- Atelier checks your model, then prints a local address with a login token. Open it in your browser, or add
-- --open. Keep that address private. - Choose a workspace folder. Atelier reads and writes files only there, and asks before running commands.
- Pick Vitruvius and ask a design question, for example: “Sketch a plan for a 4 by 5 m home studio with a desk and a bookshelf.”
How you know it worked: the app opens at its local address, accepts the workspace, shows Vitruvius and answers.
Model live in IngeTrazo
- 01
Install IngeTrazo
Install IngeTrazo, the open-source push/pull architecture modeller, yourself, and turn on its local AI bridge. Atelier never bundles it.
- 02
Open a Vitruvius chat
When a new chat opens, Atelier looks for IngeTrazo on your computer. If it answers, IngeTrazo’s own modelling tools appear in that chat; if not, Vitruvius explains how to turn it on and keeps working on the canvas.
- 03
The native panel
A panel that puts Atelier beside the model inside IngeTrazo is a separate GPL-3.0-or-later download, with its own instructions.
Images for moodboards
Without an image service, moodboards still work as colour and material cards. To add images, run your own image service behind an OpenAI-compatible images endpoint and set two values, then restart Atelier.
ATELIER_IMAGE_ENDPOINT=http://localhost:8080/v1/images/generations
ATELIER_IMAGE_MODEL=your-image-modelImage models have their own licences. Check them before you use the pictures in client work.
Plain fixes
- No model endpoint is set yet. Set
ATELIER_BASE_URLandATELIER_MODEL, or editsettings.yamlin your home folder. - Could not reach the model. Start your model server first, and check the address ends in
/v1. - The endpoint serves a different model. Copy one of the listed ids into
settings.yamlandcordis.patch.yml. - The key was refused. Set
ATELIER_API_KEY. - The port is in use. Start with
ATELIER_PORT=3082.
Honest limits
Atelier sends no telemetry. It talks only to the model and image services you set up, to IngeTrazo on your own computer, and to web pages you ask it to read.
Concept plans and moodboards are schematic design aids: check dimensions and clearances as you would a junior’s drawing. Vitruvius never certifies code compliance; structure, life safety, accessibility and permits need qualified local review. How well it works depends on the model you connect.