Atelier 0.1.0 (public preview)

Atelier 0.1.0 (public preview)

Atelier by ARCITECH is a design assistant for architects, interior designers and students. It is powered by DSH (DeepSeek Harness) 0.1.2-alpha.1 by DeepSeek-AI (MIT). This first public preview is now available, free to use: expect rough edges, and please tell us what breaks.

Download

File	Size	SHA-256
atelier-preview-0.1.0-linux-x64.tar.gz	92.6 MB	69b3f8bc53ae87ced739ffbd597b522960f288bfb7d6f07889954da47fd242fc
Atelier-Preview-0.1.0-Setup.exe	47.6 MB	b14d28391c16b539866ece7cacefffa6ea466136573304cbf5faf9f1af7ffa5a

• Windows 10/11 (64-bit): Atelier-Preview-0.1.0-Setup.exe. Installs for your user only (no administrator rights), adds a Start menu entry and an optional desktop icon, and brings its own Node.js runtime.
• Linux x64 and WSL2 (Ubuntu 22.04/24.04): atelier-preview-0.1.0-linux-x64.tar.gz. Unpack it and run ./install.sh. Also brings its own Node.js runtime.

Check a download with SHA256SUMS (Windows PowerShell: Get-FileHash <file>; Linux: sha256sum -c SHA256SUMS --ignore-missing). The files are not code-signed yet, so Windows SmartScreen may warn the first time: choose "More info", then "Run anyway".

Installation steps: docs/INSTALL.md. Connecting a model: docs/add-your-endpoint.md.

What is in it

• Free to use. Use free AI services, or run fully local on your own computer: see the free models guide.
• Starts on as little as one 16 GB graphics card. On one card the models take turns: each loads when its job comes up, and Atelier waits for it and shows what is loading. More cards, a bigger card or a model server keep them loaded side by side, so everything is faster. No GPU? Point the roles at a cloud endpoint. See One graphics card, several models.
• Mission Control. A chip in the session header shows which model is loaded, graphics card use, tokens and speed, and the energy cost at your own electricity price: see Mission Control.
• Brains. Settings > General shows the model that is really loaded now on each computer and graphics card, with Load, Unload and Reload (one brain at a time per card, with a confirm step and progress): see Brains.
• Vitruvius and Vitruvius Zero (for small models), a design canvas with concept plans and material/colour moodboards, and an optional bridge to IngeTrazo (installed separately, GPL-3.0-or-later).
• The stock DeepSeek Harness agent presets (Standard, PTC, Minimal and Creator modes) and launcher profiles.
• Works with any OpenAI-compatible model endpoint: Ollama, LM Studio, llama.cpp, vLLM, a custom URL with a key, cloud free tiers through OmniRoute, or your ChatGPT plan.
• Bottega (team mode): share one Atelier with colleagues on your office network. Other devices open it in their browser, tablet or phone after a one-time pairing code; you see and revoke every device. Off by default, private networks only, plain HTTP. Tailscale support is planned and coming soon.
• Atelier guardrails (Settings > Safety), on by default: chats read and write files only in their project folder and ask before anything wider. Turning them off (after a warning, or atelier settings set guardrails off --yes-i-understand) gives the model full file access, skips approval prompts and lifts the time limits on long waits; a Guardrails off chip stays in the session header until you turn them back on. Bottega pairing and no telemetry stay as they are.
• Settings > Updates: the version, credits, "Check for updates", and opt-in automatic updates (off by default). Updates install beside the current version, which is kept for one-click rollback.

Privacy

Atelier sends no telemetry. It talks only to the model endpoint you choose, to IngeTrazo on your computer, and to GitHub when you check for updates (or if you turn automatic updates on).

Known limits

• Tested on Windows 11 and on Ubuntu 24.04 (desktop and WSL2). Windows 10 and Ubuntu 22.04 should work but are less tested.
• Atelier ships no model. How well Vitruvius works depends on the model you connect.
• Plans, moodboards and images are design aids, not engineering or code compliance.
• Not in this preview: the drawing-first workflow, the drawing checker, wind studies, presentation boards (brief-to-board) and the tablet app.
• atelier roles test, atelier doctor and the Model roles Test button now send one tiny chat message, so "connected" means the model really answers.

Credits

Atelier is built on DSH (DeepSeek Harness) by DeepSeek-AI (MIT). IngeTrazo is by Marco Sumari Tellez and contributors (GPL-3.0-or-later). Atelier's own parts are source-available under the PolyForm Shield License 1.0.0. Website: arcitech.info.

Feedback

Feedback goes through the feedback form (no account needed) or GitHub Issues, rather than email. Never paste API keys, tokens or client material into a report.