Scan it. Then talk to it.

RakuAI turns anything your camera sees into a talkable scene — a Gaussian-splat capture, a live viewer, and an MCP surface your AI assistant can walk through, measure, and change.

Live splat viewer Drag to look around a real scanned space. A real, draggable Gaussian-splat capture mounts here. We are not faking a viewer — the Spark renderer is wired in, and a sample scan loads by default (labelled as a sample). Point data-splat-url at your own captured .spz and it replaces the sample, or open ?splat=<url> to view a shared capture.

Real today: phone scan in, draggable splat out, a live Spark render of a real sample scan on this page, share + embed below, MCP runtime live. The sample is clearly labelled so it is never mistaken for your own capture.

QR code that opens the Raku Capture web app at rakuai.com/capture-app/

Scan with your phone to start

Capture is a phone experience. Point your camera at this QR code to open the Raku Capture web app, then scan a real space.

Or open it directly: rakuai.com/capture-app/

Scan my own space Connect RakuAI to Claude

Capture. Reconstruct. Inhabit.

Three jobs, one pipeline. Reality goes in as video. A spatial world comes out. Then your AI assistant lives in it over MCP — it reads the scene, never silently rewrites your room. Reconstruction is phone/prosumer-grade today, not survey-grade.

Step 01

Capture

Point a camera at a space and move around it. A phone is enough. No rig, no markers, no booth. This is where we lead, because this is the part that has to be honest before anything else is interesting.

Step 02

Reconstruct

The frames become a Gaussian splat with real geometry and scale — a scene you can orbit, measure, and anchor to. The runtime keeps authority over physics, collision, and transforms. Captures ship as SPZ, the open splat format Niantic created and the wider ecosystem is adopting, so your scans load in ecosystem viewers and ecosystem splats load in ours.

Step 03

Inhabit

Your AI assistant connects over the same MCP contract the engine already exposes — 17 native tools (5 read-only + 12 mutation) plus universal api_call/api_search passthrough. It can see the scene, measure distances, simulate changes, and act — within permissions you set, every call audited. MCP is on every tier — it is not paywalled.

Connect your AI assistant in 2 minutes

Any model that speaks Model Context Protocol can drive a talkable scene. Works with Claude, ChatGPT, Gemini, and GitHub Copilot today — on every tier, no paywall, no extra setup beyond pasting one config block.

The MCP server exposes 17 native tools (5 read-only + 12 mutation) plus universal api_call/api_search passthrough. The core scene tools are: load_world_model, ingest_frame, get_scene_state, set_render_target, start_simulation, and get_metrics. Your assistant calls them to load a captured scene, ingest live frames, read geometry and state, and run simulations — every call is logged and deny-by-default. You stay in control.

Claude (Anthropic) ChatGPT (OpenAI) Gemini (Google) Copilot (GitHub)

Claude Desktop setup guide → MCP server docs →

What you do with a world you can talk to

Honest framing: the hero use case is in early access now. The others are where this goes — labeled so you know what is real today versus what is on the roadmap. Pro Beta is free during early access ($14.99/mo is the stated GA intent, not a live charge today).

Hero · in early access

A scan-your-space assistant

Scan a room. Then ask your assistant questions about the actual space in front of you — "will this couch fit along that wall," "how far is the window from the desk." It answers against the measured scene, not a guess.

Near-term

A game part, honestly

A captured space can become a level, a backdrop, or a playable set piece inside the RakuAI runtime. We are not promising prompt-to-game. We are saying a real scan is a real, reusable part of one.

Later phase

A smart-glasses twin

On glasses, the capture becomes a persistent spatial twin your assistant shares with you — anchored, sub-millimeter precise, built for the thermal envelope the runtime was designed around.

Later phase

Real-time vision

The longer arc is continuous capture: the world updates as you move, and the assistant reasons about it live through ingest_frame. Today that path is proven with stubs. We will say so until it ships end-to-end.

Where capture goes to work first

The loudest demand signal in the splat ecosystem right now comes from working spaces: plants, sites, and infrastructure that need a current, measurable digital twin. Our vertical wedges already point the same way, and the pipeline above is the same one underneath all of them.

Industrial environments

Walk a floor with a phone and the equipment bays, line layouts, and storage aisles become a measurable scene. An assistant can then answer the questions that used to need another site visit: clearances, reach, what fits where.

Urban infrastructure

Bridges, substations, stairwells, utility rooms. A scan turns a site walk into a scene the office can interrogate later, instead of a folder of photos someone has to interpret from memory.

Claims and inventories

A room scan is a contents inventory waiting to be read. Our insurance wedge is built on exactly this capture path: scan, reconstruct, let the assistant enumerate and estimate.

Robot memory

Robots need persistent spatial memory of the spaces they work in. The same captured scene your assistant reads over MCP is the map a robot can query. That is the premise of our robotics wedge.

The two demos we are building toward

Concept media below. These are the demos on the queue, described plainly — placeholders, not staged footage dressed up as a product.

[ media placeholder — 20-second phone → splat clip ]

20 seconds, phone to splat

A single uncut clip: someone walks a phone around a room for about twenty seconds, and a draggable Gaussian splat of that room appears. No edit cuts hiding the work. Capture is the claim, so capture is what we show.

[ media placeholder — "talk to your actual room" assistant demo ]

Talk to your actual room

The same scan, now connected to an assistant over MCP. We ask it to measure a gap, check a fit, and describe what is on a shelf — and it answers from the captured scene. The point is the assistant reasoning over your space, not a stock one.

Start scanning. Start asking.

Capture is in early access. The runtime your AI drives is not — the MCP server is live today. Scan a space with your phone, connect any MCP-capable assistant (Claude, ChatGPT, Gemini, or Copilot), and start asking real questions about a real space. Pro Beta is free during early access.

Scan my own space Connect to Claude →