**LiteReality-Agent** takes an iPhone room scan and returns a finished, **interactable** 3D scene — not a point cloud, not a single fused mesh, but a room of separate articulated objects you can open, move and light. The repository has collected **345 stars and 33 forks** since its **1 August 2026** creation and was last pushed on **17 August 2026**. It comes out of an academic group — the citation lists **Zhening Huang, Yueyan Li, Johnathan Chiu, Xiaoyang Lyu, Matt Zhou, Yuxin Yao, Joan Lasenby and Shangzhe Wu** — with a technical report still pending. Capture is the deliberately boring part. A free **LiteReality Scanner** iOS app on the App Store records one walkthrough, producing the RGB frames, depth, and the Apple **RoomPlan `room.usdz`** the pipeline consumes. That choice matters: it means the input hardware is a phone people already own rather than a rig, and the project ships an `example-scans` repository so you can run the whole thing before ever scanning a room yourself. The reconstruction splits into two stages with genuinely different characters. **Scene init is deterministic** — it turns the capture into a seed room. **Authoring is agentic**: the agent looks at the seed room, compares it against the original capture, and edits until the two match. Either half can be run alone, and a `--polish` flag layers on object refinement, materials and a model-driven quality pass. A `--live` viewer streams the build in real time alongside the agent's trace, starting before the room exists and waiting for it, so it works on a scene's very first authoring run. The dependency list is where the design becomes clear, because the agent is a **required runtime component, not an optional assistant**. You need a logged-in agent CLI on your `PATH` — **Claude Code by default, OpenAI Codex selectable via `LR_AGENT_PROVIDER`** — and that CLI drives all the reasoning. Around it sit **Blender 5.x** (tested on 5.1), the **`uv`** package manager, an image-generation API key for reference images (`OPENAI_API_KEY` by default, or Gemini via `LR_IMAGE_PROVIDER=gemini`) costing **typically under $1 per scene**, and somewhere to run **TRELLIS and GroundingDINO**. That last piece is the one real setup decision: hosted on **Modal**, whose free tier covers the workload and lets detection fan out across containers, or a local **Linux box with a 24 GB+ NVIDIA GPU**. The Modal path is what makes an Apple Silicon Mac with no discrete GPU sufficient. Output lands in formats you can take elsewhere rather than a bespoke viewer: **`room_preview/Room.glb`** with materials baked and clips intact for Blender, Unity, Unreal or the web, **`room_preview/Room.blend`** as a native Blender scene, and a `room/` directory that defines the room itself. The repository is disciplined about engineering too — an `ARCHITECTURE.md` documenting package ownership, dependency direction and the agent tool registry, plus a `sanity.py` readiness check and ruff/pytest gates, with `blender`, `scan` and `live` tests excluded from the default run. The constraints are real. It is **tested only on macOS Apple Silicon and Linux**, with no Windows path stated; it depends on Apple RoomPlan, which ties capture to recent iPhones and iPads; and every reconstruction consumes agent tokens plus image-generation credits, so cost scales with the number of rooms rather than sitting at zero. Licensed **Apache-2.0**.