Atelier: a local AI harness for architecture and design work (built on DeepSeek Harness). Private until release.
JavaScript
3
0 commits
updated Oct 12, 2026
A design assistant for architects, interior designers and students, on your own computer. Atelier by ARCITECH — powered by DSH (DeepSeek Harness).
You describe the project; Atelier works through it with you, one visible step at a time. You can stop it, change the plan or take over at any time.
Atelier 0.1.0 is a free public preview, available today. Download it below for Windows, Linux or WSL. Expect rough edges, and please tell us what breaks through the feedback form. More about ARCITECH: arcitech.info.
tar -xzf atelier-preview-0.1.0-linux-x64.tar.gz && cd atelier-preview-0.1.0-linux-x64 && ./install.sh. On WSL, copy the archive from your Windows Downloads folder first: see the Linux and WSL steps.Step by step: Installing Atelier · Add your endpoint · FAQ · Troubleshooting. Updates are in Settings > Updates: you check when you like (automatic updates are off unless you turn them on), and every update can be rolled back.
@deepseek-ai/dsh-* names so that following DSH stays easy.docs/assets/ait-mark.svg, and the repository social preview (1280×640) and avatar (500×500) are in docs/assets/brand/. Their use is covered by TRADEMARK.md.Atelier is a fork of DeepSeek Harness by DeepSeek-AI (MIT licence). We added Vitruvius, a design-assistant profile built on the classic triad Venustas (beauty), Firmitas (soundness) and Utilitas (usefulness), a design canvas with material and colour moodboards, and an optional bridge to IngeTrazo for live modelling.
New and optional: dictation, "Read aloud" and meaning-based search over your project folder, each switched on when you connect the model for it (see below).
Not in this preview: the drawing-first workflow (drawing on layers, checking and then building from the drawings), the drawing checker, wind studies, presentation boards (the full brief-to-board workflow) and the tablet app. The canvas in this preview makes schematic concept plans and moodboards, not checked drawings. The rest follows as each piece is proven.
Atelier does not include or require any particular model. It uses a few models, each for one job, and you choose every one of them under Settings > Model roles:
| Role | What it does | Needed? |
|---|---|---|
| Brain | Talks with you, plans and uses the tools. Vision-capable is best. | Yes |
| Vision | Looks at images, only if your brain cannot. | Optional |
| Speech to text / Text to speech | Dictation and "Read aloud". | Optional |
| Embeddings | Finds the right notes and documents in your project by meaning. | Optional |
| Image generation | Concept images for moodboards (for example through ComfyUI). | Optional |
| Small helper | A fast small model for quick side jobs. | Optional |
Each role takes any OpenAI-compatible address (Ollama, LM Studio, llama.cpp, vLLM, OmniRoute for cloud models, whisper.cpp, Kokoro and others), shows whether it is connected, and has a Test button. Keys stay on your computer. A role you leave empty just switches its feature off.
No model server yet? Set me up downloads a small working set (official llama.cpp servers, a Qwen3-VL brain sized for your graphics card and EmbeddingGemma 2, from 1.9 GB) into a folder you choose, after showing the sizes and checking your disk space, verifies every file and connects the roles. It is opt-in and nothing downloads until you press the button. Atelier is tuned for 27B-class vision models such as Qwen3.8-27B (Apache-2.0), but every model works through the same path. Each model keeps its own licence; check it fits how you work. Full guide: docs/ROLES_SETUP.md.
For developers. You need Node.js 24 (the tested version is in .node-version); corepack enable provides the pinned pnpm 10.34.6.
pnpm install --frozen-lockfile
# Point Atelier at your model (first run only; local servers need no API key):
ATELIER_BASE_URL=http://localhost:8000/v1 ATELIER_MODEL=your-model-id pnpm start
On Windows PowerShell, set the variables first: $env:ATELIER_BASE_URL="http://localhost:8000/v1"; $env:ATELIER_MODEL="your-model-id"; pnpm start.
Atelier checks that your endpoint answers, then prints a local address. Open it in your browser (add -- --open to pnpm start to open it for you), choose a workspace folder, and ask Vitruvius a design question. Full details, settings and the optional extras are in QUICKSTART.md, and the step-by-step guides (models, IngeTrazo, saving your work, exporting to Unreal Engine, skills, troubleshooting) are in docs/.
Prefer the keyboard? atelier opens a terminal UI on the same engine as the web app, so your chats, profiles and model roles are shared with the browser. The terminal UI has a chats sidebar, streaming answers, collapsible tool steps and a / command palette. Other commands:
atelier chat "..." answers one prompt; add --json for scripts.atelier sessions list|open|export lists, opens and saves your chats.atelier roles list|set|test shows, changes and checks your model roles.atelier profile use vitruvius|vitruvius-zero|cloud chooses the profile new chats start in.atelier settings set guardrails on|off switches Atelier guardrails (turning them off asks for --yes-i-understand).atelier update check|apply|rollback updates Atelier or goes back a version.atelier doctor checks your setup and tells you how to fix problems.The installer puts atelier on your PATH (Windows, Linux and WSL). From a checkout, run pnpm exec atelier. See docs/CLI.md for the keys, the commands and scripting examples. Atelier is made by ARCITECH.
Bottega (the Renaissance workshop where a master works with assistants) lets other people on your office network use one Atelier from their own browser, tablet or phone. One computer hosts it (the model, the tools and the project files); everyone else pairs once with a six-digit code. It is off by default; switch Share on my network on in the Bottega card on the new-session page, or start with atelier web --bottega. Only private network addresses are served, nothing is reachable before pairing, you see and revoke every paired device, and Atelier itself keeps listening on localhost only. Windows asks once whether to allow Atelier on Private networks. Anyone you pair can use Atelier as you do, so pair people you trust. Tailscale support (reaching it from outside your network) is planned and coming soon. See docs/BOTTEGA.md.
| Part | Where it runs | Needed? | Notes |
|---|---|---|---|
| Atelier | Your computer | Yes | Binds to localhost only. Bottega, when you switch it on, adds a pairing-protected door on your private network. |
| Language model | Your computer, or any OpenAI-compatible endpoint | Yes, your choice | Not included. If you use a remote endpoint, what you send goes there. |
| Image endpoint (e.g. ComfyUI) | Your choice | Optional | Concept images only; not included. |
| IngeTrazo + its official MCP bridge | Your computer | Optional | Separate install, GPL-3.0-or-later. Never bundled. |
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 and runs commands without asking; a Guardrails off chip stays in the session header until you turn them back on. Bottega pairing and the no-telemetry settings stay as they are either way. See the FAQ.
Atelier sends no telemetry. The upstream DeepSeek Harness telemetry and session-title plugins are switched off in the shipped defaults, and the launcher sets DSH_TELEMETRY_DISABLED=1. Atelier talks only to the model and image endpoints you configure, to IngeTrazo on your own computer, and to web pages the agent fetches when you ask it to. The stock web-search tool calls DeepSeek's search API only if you set your own DEEPSEEK_API_KEY. Atelier contacts GitHub only when you press Check for updates in Settings > Updates, or if you turn automatic updates on (off by default).
Atelier's own additions are licensed under the PolyForm Shield License 1.0.0 (source-available: free to use, change and share, but not to build a competing product). DeepSeek Harness under harness/ stays MIT, and ingetrazo-plugin/ stays GPL-3.0-or-later. Third-party notices are in NOTICE.md and licenses/.
| Project | By | Licence | How Atelier uses it |
|---|---|---|---|
| DeepSeek Harness | DeepSeek-AI | MIT | Atelier is a fork of it |
| IngeTrazo | Marco Sumari Tellez and contributors | GPL-3.0-or-later | Installed separately; Atelier starts its official MCP server only if you have installed it. Never bundled. |
| npm dependencies | Their authors | See licenses/ | Installed by pnpm |
The optional IngeTrazo panel in ingetrazo-plugin/atelier_panel is a separate GPL-3.0-or-later download with its own README and LICENSE.
Issues and ideas: GitHub Issues, or the feedback form below. News about ARCITECH and Atelier: arcitech.info.
Something broke, confused you, or is missing? Tell us. Feedback goes through the form or GitHub Issues rather than email, so it stays in one place and nothing gets lost.
Never paste API keys, tokens or client material into a report.
Atelier: a local AI harness for architecture and design work (built on DeepSeek Harness). Private until release.
JavaScript
3
0 commits
updated Oct 12, 2026
A design assistant for architects, interior designers and students, on your own computer. Atelier by ARCITECH — powered by DSH (DeepSeek Harness).
You describe the project; Atelier works through it with you, one visible step at a time. You can stop it, change the plan or take over at any time.
Atelier 0.1.0 is a free public preview, available today. Download it below for Windows, Linux or WSL. Expect rough edges, and please tell us what breaks through the feedback form. More about ARCITECH: arcitech.info.
tar -xzf atelier-preview-0.1.0-linux-x64.tar.gz && cd atelier-preview-0.1.0-linux-x64 && ./install.sh. On WSL, copy the archive from your Windows Downloads folder first: see the Linux and WSL steps.Step by step: Installing Atelier · Add your endpoint · FAQ · Troubleshooting. Updates are in Settings > Updates: you check when you like (automatic updates are off unless you turn them on), and every update can be rolled back.
@deepseek-ai/dsh-* names so that following DSH stays easy.docs/assets/ait-mark.svg, and the repository social preview (1280×640) and avatar (500×500) are in docs/assets/brand/. Their use is covered by TRADEMARK.md.Atelier is a fork of DeepSeek Harness by DeepSeek-AI (MIT licence). We added Vitruvius, a design-assistant profile built on the classic triad Venustas (beauty), Firmitas (soundness) and Utilitas (usefulness), a design canvas with material and colour moodboards, and an optional bridge to IngeTrazo for live modelling.
New and optional: dictation, "Read aloud" and meaning-based search over your project folder, each switched on when you connect the model for it (see below).
Not in this preview: the drawing-first workflow (drawing on layers, checking and then building from the drawings), the drawing checker, wind studies, presentation boards (the full brief-to-board workflow) and the tablet app. The canvas in this preview makes schematic concept plans and moodboards, not checked drawings. The rest follows as each piece is proven.
Atelier does not include or require any particular model. It uses a few models, each for one job, and you choose every one of them under Settings > Model roles:
| Role | What it does | Needed? |
|---|---|---|
| Brain | Talks with you, plans and uses the tools. Vision-capable is best. | Yes |
| Vision | Looks at images, only if your brain cannot. | Optional |
| Speech to text / Text to speech | Dictation and "Read aloud". | Optional |
| Embeddings | Finds the right notes and documents in your project by meaning. | Optional |
| Image generation | Concept images for moodboards (for example through ComfyUI). | Optional |
| Small helper | A fast small model for quick side jobs. | Optional |
Each role takes any OpenAI-compatible address (Ollama, LM Studio, llama.cpp, vLLM, OmniRoute for cloud models, whisper.cpp, Kokoro and others), shows whether it is connected, and has a Test button. Keys stay on your computer. A role you leave empty just switches its feature off.
No model server yet? Set me up downloads a small working set (official llama.cpp servers, a Qwen3-VL brain sized for your graphics card and EmbeddingGemma 2, from 1.9 GB) into a folder you choose, after showing the sizes and checking your disk space, verifies every file and connects the roles. It is opt-in and nothing downloads until you press the button. Atelier is tuned for 27B-class vision models such as Qwen3.8-27B (Apache-2.0), but every model works through the same path. Each model keeps its own licence; check it fits how you work. Full guide: docs/ROLES_SETUP.md.
For developers. You need Node.js 24 (the tested version is in .node-version); corepack enable provides the pinned pnpm 10.34.6.
pnpm install --frozen-lockfile
# Point Atelier at your model (first run only; local servers need no API key):
ATELIER_BASE_URL=http://localhost:8000/v1 ATELIER_MODEL=your-model-id pnpm start
On Windows PowerShell, set the variables first: $env:ATELIER_BASE_URL="http://localhost:8000/v1"; $env:ATELIER_MODEL="your-model-id"; pnpm start.
Atelier checks that your endpoint answers, then prints a local address. Open it in your browser (add -- --open to pnpm start to open it for you), choose a workspace folder, and ask Vitruvius a design question. Full details, settings and the optional extras are in QUICKSTART.md, and the step-by-step guides (models, IngeTrazo, saving your work, exporting to Unreal Engine, skills, troubleshooting) are in docs/.
Prefer the keyboard? atelier opens a terminal UI on the same engine as the web app, so your chats, profiles and model roles are shared with the browser. The terminal UI has a chats sidebar, streaming answers, collapsible tool steps and a / command palette. Other commands:
atelier chat "..." answers one prompt; add --json for scripts.atelier sessions list|open|export lists, opens and saves your chats.atelier roles list|set|test shows, changes and checks your model roles.atelier profile use vitruvius|vitruvius-zero|cloud chooses the profile new chats start in.atelier settings set guardrails on|off switches Atelier guardrails (turning them off asks for --yes-i-understand).atelier update check|apply|rollback updates Atelier or goes back a version.atelier doctor checks your setup and tells you how to fix problems.The installer puts atelier on your PATH (Windows, Linux and WSL). From a checkout, run pnpm exec atelier. See docs/CLI.md for the keys, the commands and scripting examples. Atelier is made by ARCITECH.
Bottega (the Renaissance workshop where a master works with assistants) lets other people on your office network use one Atelier from their own browser, tablet or phone. One computer hosts it (the model, the tools and the project files); everyone else pairs once with a six-digit code. It is off by default; switch Share on my network on in the Bottega card on the new-session page, or start with atelier web --bottega. Only private network addresses are served, nothing is reachable before pairing, you see and revoke every paired device, and Atelier itself keeps listening on localhost only. Windows asks once whether to allow Atelier on Private networks. Anyone you pair can use Atelier as you do, so pair people you trust. Tailscale support (reaching it from outside your network) is planned and coming soon. See docs/BOTTEGA.md.
| Part | Where it runs | Needed? | Notes |
|---|---|---|---|
| Atelier | Your computer | Yes | Binds to localhost only. Bottega, when you switch it on, adds a pairing-protected door on your private network. |
| Language model | Your computer, or any OpenAI-compatible endpoint | Yes, your choice | Not included. If you use a remote endpoint, what you send goes there. |
| Image endpoint (e.g. ComfyUI) | Your choice | Optional | Concept images only; not included. |
| IngeTrazo + its official MCP bridge | Your computer | Optional | Separate install, GPL-3.0-or-later. Never bundled. |
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 and runs commands without asking; a Guardrails off chip stays in the session header until you turn them back on. Bottega pairing and the no-telemetry settings stay as they are either way. See the FAQ.
Atelier sends no telemetry. The upstream DeepSeek Harness telemetry and session-title plugins are switched off in the shipped defaults, and the launcher sets DSH_TELEMETRY_DISABLED=1. Atelier talks only to the model and image endpoints you configure, to IngeTrazo on your own computer, and to web pages the agent fetches when you ask it to. The stock web-search tool calls DeepSeek's search API only if you set your own DEEPSEEK_API_KEY. Atelier contacts GitHub only when you press Check for updates in Settings > Updates, or if you turn automatic updates on (off by default).
Atelier's own additions are licensed under the PolyForm Shield License 1.0.0 (source-available: free to use, change and share, but not to build a competing product). DeepSeek Harness under harness/ stays MIT, and ingetrazo-plugin/ stays GPL-3.0-or-later. Third-party notices are in NOTICE.md and licenses/.
| Project | By | Licence | How Atelier uses it |
|---|---|---|---|
| DeepSeek Harness | DeepSeek-AI | MIT | Atelier is a fork of it |
| IngeTrazo | Marco Sumari Tellez and contributors | GPL-3.0-or-later | Installed separately; Atelier starts its official MCP server only if you have installed it. Never bundled. |
| npm dependencies | Their authors | See licenses/ | Installed by pnpm |
The optional IngeTrazo panel in ingetrazo-plugin/atelier_panel is a separate GPL-3.0-or-later download with its own README and LICENSE.
Issues and ideas: GitHub Issues, or the feedback form below. News about ARCITECH and Atelier: arcitech.info.
Something broke, confused you, or is missing? Tell us. Feedback goes through the form or GitHub Issues rather than email, so it stays in one place and nothing gets lost.
Never paste API keys, tokens or client material into a report.