Wadi

Wadi is a parametric home designer, and an example of an AI-native product: it is designed to be used with an AI coding assistant. You describe a house to the assistant; it writes the design in Wadi's design language (a .wdl file); Wadi compiles and renders it live as a 3D model plus dimensioned floor plans, elevations, and roof drawings. An architect can co-edit the same file, and a home-owner can personalize a finished design through a set of controls.

Web app: https://wadi.house/app · Home page: https://wadi.house

What Wadi is

Wadi turns a house into a single design you can see and shape in 3D. Every output (the 3D model, floor plans, elevations, roof drawings) is generated from the same design, so one change updates all of them.

The design is written in Wadi's design language, and it is built to be authored by an AI assistant. An architect describes the house, the assistant writes the .wdl, and both refine it while Wadi renders it live. The assistant also checks its own work: it runs an automated check after edits and renders previews it can read (see Designing with an AI assistant). The architect can edit the file by hand, and a home-owner works at a higher level, picking a ready-made design and adjusting a set of controls.

Wadi is also built to extend. Each object type is declared once in a registry file, and the schema, the documentation, the .wdl language syntax, and the decompiler all read from that one declaration, so adding a new type is a single self-contained file. The documentation/ folder explains the concepts, the personas, and the architecture. Start there if you are new.

Designing with an AI assistant

The primary way to create a design is to describe it to an AI coding assistant, which writes and checks the .wdl while Wadi renders it. You state what you want, the assistant writes the file, and you steer in chat and co-edit as needed. The assistant writes the language, not the compiled config, because Wadi's compiler does that. After each edit it runs an automated check (parse, resolve, schema, wall and roof geometry, structural conventions) and renders preview images it can read. Most functional testing today is visual; more automatic functional tests are planned.

Setup, the skill, and the tooling are under Technical details, section "The AI architect skill".

For architects (by hand)

An architect can also work without the assistant. The .wdl is the edit surface; the form-based object editor is retired. Both routes drive the same model.

• In the WDL editor built into the app, writing .wdl directly (browser: the WDL tab at https://wadi.house/app; or the desktop app, ⌘⇧D). It compiles live and shows the 3D model, per-floor plans, a filtered composite sheet, elevations, roof views, and a wall-area estimate, all dimensioned, with parse errors surfaced inline. The 🧩 Modules panel manages the reusable component files a design imports. The Wadi DSL Author's Guide is the walkthrough, from your first room to grids, reusable components, libraries, and the owner-facing configurator. It is part of the guided documentation/ folder.
• On the desktop, editing a file. The Wadi desktop app watches a .wdl or .wadi file on disk, so you can edit it in any tool and see the 3D model update within about a second. Opening a .wdl auto-loads its sibling component modules.

To make a design reusable, define it on a structural grid with formulas so it stays valid when the plot is resized, then expose a configurator. That is the basis for the templates in the home-owner gallery.

The desktop app and the AI skill need a one-time setup. See Technical details.

For home-owners

Everything runs in the browser. There is nothing to install, and it works on a phone.

1. Open the app at https://wadi.house/app (Home-owner view).
2. Choose your home. Browse the gallery of ready-made designs (Konkan cottages, family homes, and more), each with 3D previews and a floor plan. Pick the one closest to what you want.
3. Make it yours. A set of controls (plot size, roof style, room options) reshapes the house, and the 3D model updates as you change them. Orbit, zoom, and look around.
4. See the plans. Switch to the floor-plan view for room sizes and layout.
5. Share it. The Share button copies a link that contains your whole design. Send it to family or to your architect and they see what you see. No account, nothing uploaded.

Technical details

Everything below is for running, extending, or deploying Wadi. A house is authored as a .wdl (the Wadi Design Language) and saved as a .wadi bundle. One TypeScript codebase under editor/src renders every surface, and the Zod schema at editor/src/schema/houseConfig.ts is the single source of truth for the compiled config.

The .wadi bundle & schema

• A .wadi is a zip bundle that carries a design and its dependencies together:
  • wadi.json — the compiled HouseConfig (what every renderer reads).
  • model.wdl — the authored DSL source, so the design round-trips back to editable .wdl.
  • thumbnails/ — gallery preview images.
  • modules/ — the imported component .wdl files, one per import ref, mapped in the manifest, so a design that reuses custom components stays self-contained (see components & libraries).
• A plain-JSON .wadi or house_config.json still loads (the loader sniffs the magic bytes: a zip vs. a { document vs. .wdl text), so older single-document designs keep working. The repo's default design is docs/house_config.json (the root house_config.json is a symlink to it) and auto-loads in the app.
• The compiled config is defined and validated by the Zod schema (editor/src/schema/houseConfig.ts). The complete, always-current field reference is wadi-skill/architect/reference/data-model.md, generated from that schema so it cannot drift.

Each floor holds an objects list; every object has a type:

Coordinates, units & conventions

• Inkscape frame: origin top-left, X to the east, Y down (south). Z is up.
• 10 project units = 1 ft by default (an optional units block relabels the display without changing geometry).
• Centreline convention ("coord_convention": "center"): a rect object's x/y/width/length are wall centrelines, so adjacent rooms simply abut on the shared line (no overlap, no wall math). (A pillar's x/y is its top-left corner, like a room's — centre a column on a node by subtracting half its width.)
• Grids: a parametric template defines named grid lines whose positions are formulas of the plot size; rooms and columns reference them ("= main.x1"), so the whole plan re-flows when you resize.

Full details: wadi-skill/architect/reference/coordinate-system.md and parametric-conventions.md.

The AI architect skill: setup & use

Describe a house to a coding agent and it authors the design in the Wadi DSL, a small formal language saved as a .wdl file, while the Wadi WDL editor compiles and renders it live. The .wdl is a single file you and the agent co-edit; the agent never hand-writes the compiled config, because the compiler does that conversion.

The skill ships as an agent-neutral core plus thin per-agent adapters, so the same content drives any coding agent:

wadi-skill/architect/            the skill: instructions, references, examples, scripts
.claude/skills/wadi-architect/   Claude Code adapter        → points at the skill
AGENTS.md                        vendor-neutral entrypoint (Antigravity, Cursor, …) → the skill

1. Install (once)

git clone https://github.com/bijoor/wadi.git
cd wadi
npm --prefix editor   install    # the skill's check/preview scripts reuse the app's TypeScript
npm --prefix wadi-dsl install    # the DSL compiler (its `prepare` hook builds the parser)

2. Open the WDL editor for the live preview (recommended)

Keep your .wdl open in the Wadi WDL editor so the 3D model + plans update as the agent saves (and it watches the file, so your own edits show too):

• Desktop: the Wadi app → ⌘⇧D → Open your .wdl (see Desktop app).
• Browser: the WDL tab in the app at https://wadi.house/app, paste or open a .wdl.

3. Ask the agent to design

• Claude Code auto-discovers the skill from .claude/skills/ when run inside the repo: cd wadi && claude, then ask naturally or invoke /wadi-architect. It creates a .wdl, tells you the path, and runs check.sh after every edit. To use it in every project, copy the adapter to your user skills dir (it still points back at this repo, so keep it checked out): cp -R .claude/skills/wadi-architect ~/.claude/skills/.
• Antigravity / Cursor / other AGENTS.md-aware agents read the repo-root AGENTS.md, which routes to wadi-skill/architect/SKILL.md. If yours doesn't pick it up, tell it to follow that file.

4. The tooling the agent runs (you can too)

Both scripts reuse the app's own generators, so they match the app byte-for-byte. Pass an absolute path to the .wdl:

wadi-skill/architect/scripts/check.sh   <ABS_PATH_TO.wdl>   # parse + resolve + schema/geometry + conventions
wadi-skill/architect/scripts/preview.sh <ABS_PATH_TO.wdl>   # render plans / elevations / roof to PNGs

check.sh also enforces the structural conventions (plinth match, exterior walls, no-slab floors, and more), failing on errors and printing warnings. The full workflow, the conventions, and the per-agent setup are in Designing with an AI assistant.

Repo-free: run it as an MCP server

Everything above needs the repo checked out (the scripts reuse the app's TypeScript). To use the skill without a clone, run wadi-mcp/, an MCP server that bundles the whole pipeline (DSL compiler + schema + geometry + conventions + 2D renderers) and the examples/reference docs into one self-contained file. It exposes agent-native tools:

Agents can also register reusable component modules so they bundle into the .wadi; see components & libraries.

Register it with any MCP client (Claude Code, Cursor, Claude Desktop, …). Published to npm, it needs no install:

{ "mcpServers": { "wadi": { "command": "npx", "args": ["-y", "wadi-mcp"] } } }

Or from a local build (cd wadi-mcp && npm install && npm run build → dist/server.mjs, only @resvg/resvg-js external), point at the bundle by path:

{ "mcpServers": { "wadi": { "command": "node", "args": ["/abs/path/to/wadi-mcp/dist/server.mjs"] } } }

Then the agent authors a .wdl using wadi_reference / wadi_examples and verifies with wadi_check / wadi_preview, with no repo and no desktop app. (The live 3D preview still comes from the WDL editor if you want it.) See wadi-mcp/README.md.

Run & develop locally

git clone https://github.com/bijoor/wadi.git
cd wadi
npm --prefix editor install

All app code lives under editor/:

npm run build runs tsc then a Vite build of the app → docs/app/ (vite.viewer.config.ts). npm run dev / Vitest use vite.config.ts. Both read the same editor/src.

Desktop app (Tauri)

Requires Rust and the Tauri CLI (cargo install tauri-cli). From the repo root:

npm --prefix editor run build          # refresh docs/ (the app bundles it as its frontend)
cargo tauri build --bundles app        # → src-tauri/target/release/bundle/macos/Wadi.app

src-tauri/tauri.conf.json bundles docs/ as the frontend and associates the .wadi file type, so double-clicking a design opens it and the app live-watches it.

Repo layout

editor/                    React + Three.js + Zod app: THE source of truth
  src/schema/houseConfig.ts   the .wadi schema (Zod)
  src/{svg2d,three,param}/    2D drawings, 3D scene, parametric engine
  src/registry/               object-type registry (one file per new type)
docs/                      GitHub Pages root, served at wadi.house
  index.html               landing page  → /app
  app/                     the app (viewer): 3D model + plans + WDL editor + configurator
  templates/               bundled starter templates (self-describing; the folder is the catalog)
  house_config.json        default design (root house_config.json → symlink)
src-tauri/                 desktop app (Tauri) wrapping docs/
wadi-dsl/                  the Wadi DSL (.wdl): Langium grammar, compiler, decompiler
  examples/*.wdl              validated sample houses the skill copies from
wadi-skill/architect/      agent-neutral AI skill (Claude Code, Antigravity, …)
wadi-mcp/                  MCP server: the skill's tooling, repo-free (check/preview/examples)
.claude/skills/            Claude Code skill adapters
AGENTS.md                  vendor-neutral entrypoint for coding agents
library/                   reference parametric .wadi models
scripts/                   publish-templates.sh / publish-furniture.sh (Cloudflare R2)
schema/                    legacy JSON Schema (superseded by the Zod schema)
python/  archive/          RETIRED Blender/Python pipeline: history only, do not use

The original Blender/Python generator has been retired. Everything now runs in TypeScript. python/ and archive/ are kept for history only.

Deployment & hosting

• Site + app: docs/ is the GitHub Pages root on the custom domain wadi.house. Run npm --prefix editor run build so docs/app is current, then push main. Pages rebuilds in about 1 to 2 minutes.
• Templates & furniture: hosted on Cloudflare R2 so new ones ship without redeploying the site or app. Publish with scripts/publish-templates.sh / scripts/publish-furniture.sh (needs a gitignored .env.r2; see TEMPLATE_HOSTING.md).

References

• documentation/: the guided documentation. Start at documentation/README.md. It walks a newcomer from the concept and the personas (AI coding assistant, architect, home-owner, developer) through the authoring guide and components & libraries, to the advanced extending-the-DSL and generalizable-method chapters.
• wadi-skill/architect/: the AI skill. SKILL.md (the step-by-step authoring loop) + reference/{dsl,data-model,coordinate-system,parametric-conventions,conventions, roof-v2-guide}.md + validated examples/. The reference docs are the authoritative, current description of the DSL, the model, and the structural conventions.
• wadi-dsl/README.md: the Wadi DSL (.wdl): grammar, compiler, and decompiler.
• editor/README.md: editor internals and architecture.
• TEMPLATE_HOSTING.md: R2 bucket + CORS setup for templates / furniture.

License / attribution

Personal project for the Aatley Home Construction site. The app scaffolding (TypeScript / React / Three.js / Tauri) is MIT-style: reuse freely, no warranty. The house designs themselves are not licensed for reuse.