Sozai Search MCP

English | 日本語 | 简体中文 | 한국어 | Português | Español

Quick start · Tools · Sources · Semantic search · FAQ · ⭐ Star

Sozai Search MCP gives Claude Code, Claude Desktop, Cursor and any other MCP client a library of 36,000 free 3D assets: photoreal models, PBR materials and HDRIs from Poly Haven and ambientCG, low-poly game models from Kenney and Quaternius, and Quixel Megascans listings. Ask for "a wooden crate under 1 m tall" or "a mossy rock"; the agent searches by words or by meaning, looks at the previews, and drops the files into your project, at their real size in metres. (Sozai, 素材, means "materials" in Japanese.)

Sozai Search is an index. It hosts no asset files: every result points at the official source, and files are downloaded from there by your own machine.

✨ Highlights

• Looks before it picks: view_assets shows the agent up to 8 previews, so it chooses by eye, not by title.

• Arrives at the right size: kit models are calibrated to metres (a door is 2.1 m, a person 1.75 m) and delivered already scaled; rigs and animations keep working.

• Light: the server is standard-library Python with no runtime dependencies. It starts in about 25 ms, idles at about 23 MB and holds no state between calls.

• Safe with untrusted data: https only, to the libraries' own hosts; every path stays inside the folder you gave; it never deletes your files.

Related MCP server: Blendkit MCP Server

🧭 How it works

"a wooden crate under 1 m tall"
   └─ search_assets   keywords + meaning, filtered by kind, size, height, polycount, licence
   └─ view_assets     the agent looks at the candidates' previews
   └─ fetch_asset     downloads from Poly Haven / ambientCG (or your local kit zip),
                      packs one .glb at real scale, appends the credit to CREDITS.md

sozai build downloads the published data once: one zip with every library's records (what each asset is, its size, licence and where to get it) and the search vectors. No asset files, no previews.

🚀 Quick start

🤖 Install with your AI agent

Give this repository's URL to Claude Code, Codex or another agent and ask it to install Sozai Search MCP. These are the steps it should follow (they work by hand too):

1. uv: check uv --version; if uv is missing, install it:

   • macOS / Linux: curl -LsSf https://astral.sh/uv/install.sh | sh

   • Windows: powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

2. Register the server: run this in the user's own project folder. No clone is needed: uv fetches Sozai Search from GitHub.

   # Claude Code: this project only (writes .mcp.json)
   claude mcp add sozai -s project -- uvx --from git+https://github.com/kanaworksai-hue/sozai-search-mcp sozai
   
   # Codex: all projects (writes ~/.codex/config.toml)
   codex mcp add sozai -- uvx --from git+https://github.com/kanaworksai-hue/sozai-search-mcp sozai

   Other clients (Claude Desktop, Cursor, ...) take the same command in their JSON config:

   {"mcpServers": {"sozai": {"command": "uvx", "args": ["--from", "git+https://github.com/kanaworksai-hue/sozai-search-mcp", "sozai"]}}}

3. Start a new session: MCP servers load when a session starts. Claude Code asks once to approve the project's .mcp.json. Check with claude mcp list or codex mcp list.

4. Try it: search for "wooden crate". The first search downloads the published data (about 13 MB) and builds the index in a few seconds. Give fetch_asset an absolute folder inside the project.

5. Search by meaning (optional): for descriptive queries in any language, download the query model once (about 350 MB):

   uvx --from "sozai-search-mcp[semantic] @ git+https://github.com/kanaworksai-hue/sozai-search-mcp" sozai semantic-setup

   Then register with --from "sozai-search-mcp[semantic] @ git+https://github.com/kanaworksai-hue/sozai-search-mcp" in place of --from git+https://github.com/kanaworksai-hue/sozai-search-mcp.

🛠️ From a clone

You need uv (it brings Python 3.10 or newer).

git clone https://github.com/kanaworksai-hue/sozai-search-mcp
cd sozai-search-mcp
uv sync
uv run sozai build            # downloads the published data (one zip, no asset files) and indexes it
uv run sozai search "wooden crate" --kind model
uv run sozai fetch polyhaven:wooden_crate_01 ./assets --quality web

Claude Code (registered for one project only):

claude mcp add sozai -s project -- uv run --directory /path/to/sozai-search-mcp sozai

Any other MCP client (Claude Desktop, Cursor, ...):

{
  "mcpServers": {
    "sozai": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/sozai-search-mcp", "sozai"]
    }
  }
}

Without a clone, uv can run it straight from GitHub: uvx --from git+https://github.com/kanaworksai-hue/sozai-search-mcp sozai. The first search builds the index if sozai build has not run yet. Give fetch_asset an absolute folder: an MCP server does not run in your project folder.

🧰 Tools

Sizes are [width, height, depth] in metres, height being the up axis. Materials give [width, height] of the area one texture tile covers.

📚 Sources and licences

Megascans are free on Fab: claim them with an Epic account, then download. You may use them in your projects but not redistribute the files. A listing's three dimensions come in no fixed axis order, so they are reported as dims_m (largest first) and kept out of height filters.

If the data release cannot be reached, sozai build reads the libraries' APIs directly: Poly Haven and ambientCG (about 5,300 assets), without kit records, Megascans or semantic search. sozai build --live does the same on purpose, for assets newer than the release. After you update Sozai Search, the first search moves to the data release the new version expects.

🔎 Semantic search (optional)

uv sync --extra semantic
uv run sozai semantic-setup   # downloads the query model once (about 350 MB)
uv run sozai build            # if not done yet: the asset vectors come with the published data

Register the server with the extra so it can load the model: uv run --directory /path/to/sozai-search-mcp --extra semantic sozai (in a JSON config, add "--extra", "semantic" before "sozai").

Search then also matches by meaning, in any language: "something to sit on in a medieval tavern" finds stools and benches. Each asset has a vector of its words and, when it has a preview, of its picture (EmbeddingGemma 2, one space for text and images), so kit models with bare names are found by what they look like. The ranking fuses the keyword, text and image rankings.

Only the query runs at search time: the text model as ONNX int8 on the CPU, about 45 ms, pinned to one revision and checked by SHA-256. It runs in a worker process that starts with the first search and quits after five idle minutes, so an idle server keeps its small footprint.

🧩 Kit libraries (Kenney, Quaternius)

These libraries have no API. Their models are listed in the published data, but fetch_asset can deliver them only after you download the kits and ingest them:

uv run sozai ingest kenney ~/Downloads/kenney-kits
uv run sozai ingest quaternius ~/Downloads/quaternius-packs

Keep Kenney's zip names as downloaded (kenney_city-kit-commercial_2.1.zip). Name Quaternius zips quaternius_<pack>.zip, where <pack> is the name in the pack page's address (quaternius.com/packs/<pack>.html); suffixes such as _fbx or _part2 are fine. Then the ingested models replace their published records instead of appearing twice.

Every model is measured (size, triangles, rig, animations) and gets the kit's preview image. FBX-only models are converted to GLB with Blender when it is installed (rigs and animation clips are kept), and models without a preview get one rendered by headless Blender. Kits with a trustworthy reference object are calibrated to metres; kits without one keep their file units in attributes.size_units and report no size, so size filters never mistake them for metres. A kit whose licence file is not CC0 is recorded with a pointer to that file and left out of redistributable_only.

⚙️ Configuration

Environment variables, all optional:

Sozai Search only connects over https, to the libraries above, GitHub (the data release) and Hugging Face (the query model).

❓ FAQ

Poly Haven, ambientCG, Kenney and Quaternius assets are CC0: free for any use, no credit required (Sozai Search still writes CREDITS.md, and Poly Haven asks for "Powered by Poly Haven" where you show their assets through the API). Megascans are under the Fab Standard License: use them in your projects, but don't redistribute the files. Every result states its licence.

Search with redistributable_only=true: it keeps CC0 assets and leaves out Megascans.

Only the owner of a Fab account can download them, after claiming them (free). Sozai Search gives the agent the Fab link to pass on to you.

No. Keyword search is plain SQLite; semantic search runs a small model on the CPU. Nothing needs an account or a key.

Kenney and Quaternius publish zips, not an API. Download the kit from the link Sozai Search gives, run sozai ingest, and fetch_asset takes the model straight from your zip.

⭐ Like it? Star it

If Sozai Search found you an asset you ended up using, please click ⭐ Star at the top of the repository page. It's free, takes one second, and helps more than it looks:

• More people find it. Stars make the project easier to discover on GitHub.

• It keeps updates coming. Every star tells me someone finds this useful.

Other ways to help: open an issue for a bug or an idea, or share what you built on X and tag @KanaWorks_AI.

🛠️ Development

uv run pytest                 # offline tests
uv run python eval/run.py     # search quality on the eval queries (needs a built index)

src/sozai/
  server.py        MCP tools and the command line
  protocol.py      MCP over stdio (standard library only)
  search.py        keyword + semantic ranking, filters
  semantic.py      query model worker (ONNX) and its client
  index.py         the SQLite index
  build.py         builds the index from the published data or the libraries' APIs
  bundle.py        the published data zip (pack / install)
  fetch.py         downloads, packing .glb, credits
  catalog.py       kit ingest (Kenney, Quaternius); scale.py calibration, thumbs.py previews
  sources/         Poly Haven, ambientCG, Megascans

Maintainers make a data release with sozai build --live, sozai ingest ..., sozai embed (MLX, uv sync --extra embed) and sozai pack.

📄 License and credits

• This project: MIT, © 東京AI映像合同会社 (Tokyo AI Visuals LLC).

• Assets: each asset keeps its own licence, given per result. Sozai Search distributes no asset files. The published data carries its own NOTICE.txt.

• Brand: the KanaWorks_AI name, logo and the Kana character are © 東京AI映像合同会社 and are not covered by the MIT license. See docs/brand.

• Built on: Poly Haven (Powered by Poly Haven), ambientCG, Kenney, Quaternius, EmbeddingGemma 2 and the Model Context Protocol.

This is an independent project. It is not affiliated with or endorsed by Epic Games, Quixel, Poly Haven, ambientCG, Kenney, Quaternius, Google or Anthropic.