Atelier
Converts spoken descriptions into standards-compliant floor plans and furnished 3D models, enabling real-time collaborative editing through a live browser editor with Claude AI.
README
English · 🇻🇳 Tiếng Việt
Atelier 📐
An AI architect that plugs into any MCP client — Claude Code, Codex CLI, Gemini CLI, Claude Desktop, Cursor… From a spoken description to standards-compliant floor plans and a furnished 3D model, with a live browser editor you can edit by hand while the AI works.
Built for the Vietnamese tube house (nhà ống): TCVN drawing symbols, Vietnamese dimensional rules, Lỗ Ban feng-shui rulers, irregular lot polygons — things no Western tool ships as first-class citizens. The editor UI is bilingual and auto-detects by country: Vietnamese in Vietnam, English elsewhere (timezone + browser language — no geo-IP; the VI/EN button overrides and is remembered). Exported drawings stay Vietnamese by design: they are TCVN documents.
Atelier /ˌætəlˈjeɪ/ — "design workshop".
| Standards-compliant plan, 3-chain dims | First-person walkthrough (WASD) | A/B design comparison |
|---|---|---|
![]() |
![]() |
![]() |
Core idea
One parametric model (JSON) is the single source of truth. Claude edits it through MCP tools; humans edit it directly in the web editor (drag, type exact numbers); both sync in real time over WebSocket. A deterministic renderer generates symbol-correct 2D drawings and the furnished 3D model from the same source — so 2D and 3D can never drift apart, and drawing symbols are always correct because the AI never draws them.
Claude Code ──MCP──► Server (model + validator + renderer) ◄──WebSocket──► Browser (live editor)
Install (30 seconds)
You don't clone anything — Atelier ships as an npm package (atelier-mcp), and it's a standard MCP server, so it is not tied to Claude: any MCP client can run it. The server is always the same one command — npx -y atelier-mcp — only the registration step differs per client.
Stand in your own house folder (one folder = one house) and run ONE command — it detects every MCP client on your machine and registers with each:
mkdir my-house && cd my-house
npx -y atelier-mcp setup # finds Claude Code / Codex CLI / Gemini CLI /
# Claude Desktop / Cursor and registers Atelier
npx playwright install chromium # optional, once — PNG/PDF export
<details> <summary>Registering manually instead (or for other MCP clients)</summary>
claude mcp add atelier -- npx -y atelier-mcp # Claude Code
codex mcp add atelier -- npx -y atelier-mcp # OpenAI Codex CLI
gemini mcp add atelier npx -y atelier-mcp # Gemini CLI
GUI clients (Claude Desktop, Cursor, …) have no "current folder" — point ATELIER_DIR at your house folder in their MCP config JSON:
{ "mcpServers": { "atelier": {
"command": "npx", "args": ["-y", "atelier-mcp"],
"env": { "ATELIER_DIR": "C:\\path\\to\\my-house" } } } }
</details>
Then start your agent and say "design me a 4×16m tube house". Works best with a model that reads Vietnamese well (tool descriptions and validator messages are Vietnamese-first) and a client that shows images from tool results — the agent inspects its own drawings.
From source (contributors)
pnpm install
npx playwright install chromium # PNG/PDF rendering + demos
pnpm build:web
pnpm demo:p5 # the 60-second demo: live build → drag a wall,
# type 4200 → walk through the house in 3D
The repo ships a ready .mcp.json — open Claude Code in this folder and it registers the atelier server from source. pnpm build:cli builds the npm package (packages/cli, single-file bundle + web editor + fonts; verify with npx tsx packages/cli/smoke.mts <installed-pkg-path>).
What's the "60-second demo"?
The project's definition of success (doc 01), and it runs end-to-end (pnpm demo:p5):
Describe "4×16m tube house, 2 floors, 3 bedrooms, parking for 2 motorbikes" → a two-storey plan with correct symbols grows live in the browser → the user grabs a wall, types
4200 ⏎→ dims update, Claude notices the change and adapts the adjacent rooms → switch to 3D and walk (WASD) through the furnished house.
Status — roadmap complete
All six phases (P0–P5) plus the first three backlog items shipped:
| Phase | Delivered | Demo |
|---|---|---|
| P1 — Engine + static drawings | Parametric model, 22-rule validator with sources checked against TCVN 13967:2024, SVG/PNG plan renderer (symbols + 3-chain dims), 7 MCP tools, design-process skill | demo:p1 |
| P2 — One-way live | editor_open launches the browser editor (Vite + Three.js) in the same process; every apply_ops appears instantly in 2D (server-rendered SVG) and 3D (incremental maquette); capture_view lets Claude see exactly what the user sees |
demo:p2 |
| P3 — Hand editing | Drag walls/doors/furniture with a type-a-number HUD (drag, type 4200 ⏎ → exactly 4200mm), snapping, per-origin undo, soft-locks ("the user always wins" — agent writes on a user-held entity are rejected), editable properties panel, get_changes_since with old → new summaries |
demo:p3 |
| P4 — Documentation set | Numbered sheet set KT-01…: floor plans (clear-dimension interior dims, section mark), front elevation, section A-A through the stair (poché, door voids, step profile), room & door schedules; export multi-page A3 PDF, DXF (pure TS, true mm — measurable in CAD), SVG | demo:p4 |
| P5 — Furniture & experience | 106-asset catalog (true mm sizes, clearances, CC0 parametric geometry + license manifest), per-category 3D builders, floor finishes, catalog panel in the editor, WASD walkthrough (pointer lock, 1600mm eye height), sun study (hour/month sliders, real shadows from latitude + site orientation), assets_search with a visual contact sheet |
demo:p5 |
Backlog shipped: rough cost estimate (estimate_cost + a KT-06 sheet: converted areas × an editable local price table, auto-checked against the interview budget) · view-only share links (/xem/<token>, server-enforced read-only — send it to your spouse) · A/B design comparison (variant_save/open/compare, two plans side by side with per-room m² and per-variant cost deltas) · local zoning rule pack (PLN-01…06: setbacks for any lot shape, density, floors, height, canopy overhang — checks only what the interview brief declares) · IFC4 export (export ifc, pure-TS STEP writer: walls with proper opening voids/fills, slabs with real holes, stairs, rooms as IfcSpace — concept-level handover, not construction documents) · photoreal path (export gltf → one GLB in true metres, one node per entity, plus scripts/render-photoreal.py for headless-Blender Cycles renders using the same sun model as the editor's sun study) · legacy plan import (underlay_import traces an old DXF or plan photo as a dim layer under the live plan — scale from DXF units or a 2-point calibration; underlay_trace proposes walls from parallel stroke pairs, always human-reviewed before applying).
Bilingual: English README + full English design docs (docs/en/) + bilingual editor UI auto-detected by country (VN → vi, elsewhere → en; asset catalog included). Out of scope by design: structure & MEP (this is a concept tool, not construction documents).
Design docs
The full spec lives in docs/en/ (English) with the Vietnamese originals in docs/ — each file cross-links its counterpart; the Vietnamese original is authoritative if they differ:
| # | Doc | Contents |
|---|---|---|
| 01 | Vision & positioning | Users, non-goals, competitive landscape, differentiators |
| 02 | Design process | Interview → brief → 5 phases with approval checkpoints |
| 03 | System architecture | Dual-client, single process, data flow |
| 04 | Data schema | Parametric model: walls, openings, rooms, stairs, furniture |
| 05 | MCP tools | 19 tools + the ops vocabulary |
| 06 | Sync protocol | WebSocket, revisions, conflicts, undo |
| 07 | Validator & rules | TCVN rules + geometry + Lỗ Ban rulers |
| 08 | 2D renderer | Layers, line weights, symbols, dims, title block |
| 09 | Web editor | Hand-editing UX: drag, type numbers, 3D, walkthrough |
| 10 | Roadmap | 6 phases with Definitions of Done + backlog |
| 11 | Decisions | 10 ADRs + 8 settled questions |
Stack
TypeScript monorepo — Node.js (MCP server + WebSocket), Three.js (3D), SVG (2D drawings), Vite (web editor), Vitest + Playwright (162 tests incl. real-browser e2e). Rationale in doc 03.
Dev environment
The repo integrates haido — a code-anchored memory layer that served as the project's own long-term memory while Atelier was being built:
.mcp.json→haido serveMCP server (recall / remember / related / overview / stale / reanchor).claude/settings.json→ hooks that feed project knowledge into every Claude Code session- Local DB in
.haido/(gitignored); shared knowledge ships as a memory pack indocs/memory-pack/
License & disclaimer
- Code: MIT. Font Be Vietnam Pro: SIL OFL 1.1 (
OFL.txtnext to the font files). Furniture catalog (parametric): CC0-1.0 (packages/core/assets/license-manifest.json). - Standards data in
rules/*.jsoncites TCVN/QCVN sources per rule; the price tablerules/don-gia.jsonis reference-only — edit it for your region. - ⚠ Atelier outputs are concept drawings — for thinking, discussing and working with an architect; they do not replace permit/construction documents signed by a licensed architect.
The 5 invariants
- Claude never draws directly — it only edits the parametric model; the renderer draws.
- Standard symbols live in the renderer — encoded once per TCVN, correct 100% of the time.
- Every mutation goes through one gate (
apply_ops) — transaction, validation, broadcast. - The user always wins — hand edits are never overwritten by the AI.
- Claude must look before inviting the user to look — validate + render + capture after every significant change.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。


