Describe a room and walk around it: the AI doesn't draw the furniture, it finds real pieces, in meters

Ask an AI for a 3D sofa. It might come out pretty nice. But how long is it? A meter? Four? The model has no idea.

Now here's the unexpected part. To build a room, you don't need to generate furniture at all. People have already scanned and modeled thousands of objects and given them away for free. Your agent just couldn't search them before.

Sozai Search came out on November 8. It's an MCP server: a small program that gives your agent a new tool. Now Claude Code or Codex can search 36,000 free 3D assets and see each one's size in meters.

The idea in one line: you describe a room in words, and then you walk around it in a browser, with a sofa of the true length.

Where the trick is

The agent works like a person with a catalog, in three steps:

• It searches. A sofa under two meters, say. Size is a filter, just like the name.
• It looks. It opens up to eight previews and chooses by eye, not by the caption.
• It fetches. It downloads a .glb file from the library's own site and adds the author to CREDITS.md.

The nicest part is the meters. Kit models are calibrated: a door is 2.1 m, a person 1.75. Poly Haven has a sofa called Sofa 02, about 1.8 m long, and it lands in your scene exactly that size.

The glTF format counts in meters by its spec. Three.js, the browser 3D library, reads it with no setup. So never scale models by eye. One scene unit is one meter.

What you'll learn

• Connecting MCP. One command and your agent has new hands. You'll use this trick everywhere.
• 3D basics. Camera, light, coordinates, loading a model.
• Measuring. A scene in meters lets you work out gaps and walkways with simple math.
• Licenses. What CC0 means, and why crediting the author is still nice.

A ready starter prompt

Connect the server in an empty project folder. If you don't have uv, the agent will install it:

claude mcp add sozai -s project -- uvx --from git+https://github.com/kanaworksai-hue/sozai-search-mcp sozai

Start a new session so the tool loads. The first search downloads the index, about 13 MB. Now the task:

The weak prompt sends the agent off to draw cubes. The strong one makes it search, look and measure. The Box3 check is math on the items' bounding boxes. Without it, a bookcase can easily end up half inside a wall.

What you get

You open the page. You're standing in the doorway at eye height. In front of you there's a wall, a sofa, a lamp in the corner. The light looks real, because a photo of a real sky is lighting the room.

You walk forward and bump into the table. You turn on the tape measure and click: 64 centimeters from table to sofa. Tight. You tell the agent to move the table to the window. A minute later you're walking around again.

Now enter the dimensions of your actual room. Measure from the door to the window. And now you're testing a rearrangement before you ever move the bookcase.

You can share it too. Put the folder on a host and send a friend the link: come see my future living room.

It doesn't have to be a home. The coffee shop of your dreams. A room from a favorite book. An office you'd actually want to work in.

Honest limits

The catalog is big, but your sofa from the store isn't in it. Poly Haven has 85 furniture models; the rest is other objects, materials and light panoramas. Need your exact sofa? Ask the agent for a box with your dimensions.

The server can't download some libraries. The low-poly Kenney and Quaternius kits you fetch by hand, and Megascans only show up as links. Plain search works in English, but the agent translates for you.

Opening the page with a double click won't work: the browser won't load the models without a local server. That's why the prompt asks for one.

The most useful AI in 3D turned out not to be the one that draws. It's the one that can search and measure.

Source: Sozai Search MCP — 36,000 free 3D assets with real-world sizes (GitHub)