AI home furnisher — scan a room with iPhone RoomPlan, the agent designs and places real furniture in a 3D reconstruction. See PRD.md for vision and docs/architecture.md for the room-→-agent layer breakdown.

• Node.js 20+ and npm
• An Anthropic API key with access to claude-sonnet-4-6
• A Composio API key (free tier is fine)
• For scanning: an iPhone/iPad with LiDAR + Xcode (only needed if you want a fresh scan; a pre-baked scan ships in scans/)

npm install

# Create .env.local at the repo root with both keys (gitignored, never committed):
cat > .env.local <<'EOF'
COMPOSIO_API_KEY=ak_...
ANTHROPIC_API_KEY=sk-ant-api03-...
EOF

That's it — the demo room (scans/room.raw.json), stub catalog (data/furniture.json), and all code are in the repo.

npm run dev

Then open http://localhost:3000/scan — that's the only page that matters right now. You'll see:

• Left: 3D room scan (orbit / press V to walk through with WASD).
• Right: 4-tab agent panel.

In the Chat tab, send something like:

Place a small reading nook with a warm-wood lounge chair and a brass floor lamp under $1500 total. Use query_space first.

Tool calls fire in real time. You'll see:

1. GET_ROOM — agent fetches the compact room JSON
2. QUERY_SPACE — finds clear / near-wall placement spots
3. SEARCH_FURNITURE — filters the catalog
4. PLACE_ITEM — succeeds (with adjustments[] if snapped) or fails (with blocking[])

The Debug tab refreshes after each turn so you can watch placements accumulate.

These are CLI scripts that hit the same code paths as the API, useful for fast iteration:

# Eyeball the room rep + token budget
npx tsx scripts/print-room-context.ts

# Snap + collision pipeline (4 seeded cases)
npx tsx scripts/dry-run-place.ts

# Full end-to-end agent turn (needs API keys exported in shell)
set -a && source .env.local && set +a
npx tsx scripts/dry-run-agent.ts "Cozy modern reading nook, $1500, warm woods"

The pre-baked scan in scans/room.raw.json is enough for development and the hackathon demo. To capture your own:

1. Open ios-scanner/RoomScannerApp.swift in Xcode.
2. Create a new iOS App project, replace the auto-generated app file with this one, set deployment target to iOS 17.
3. Add NSCameraUsageDescription to Info.plist, set your signing team, build to a LiDAR device.
4. Walk slowly around the room → tap Done → AirDrop room.usdz and room.raw.json to your Mac.
5. Drop both into scans/. Restart npm run dev.

app/
  scan/           page.tsx, scan-canvas.tsx (3D), panel.tsx (agent UI)
  api/
    chat/         POST – SSE-streamed agent loop
    agent-context/ GET – catalog + room rep + role prompt + tokens
    placements/[id]/ PATCH/DELETE – user-drag handler
lib/
  roomplan.ts     raw RoomPlan types + transform helpers
  room/           normalize, describe, serialize, segments, regions,
                  grid, snap, place, query_space
  agent/          state, catalog, tools (10 Composio tools), loop
data/
  furniture.json  stub catalog (10 items)
scans/
  room.raw.json   pre-baked demo scan
  room.usdz       USDZ fallback for AR viewers
docs/
  architecture.md layer-by-layer diagrams + plug-in points
scripts/
  print-room-context.ts, dry-run-place.ts, dry-run-agent.ts, spike-composio.ts

npx tsc --noEmit          # type-check
npx next build            # production build

Both run clean as of the last commit.

• PRD.md — full hackathon spec
• docs/architecture.md — room-→-agent layer breakdown with mermaid diagrams
• Composio docs
• Anthropic prompt caching
• Apple RoomPlan