Claude

The open-source route is the blender-mcp project by Siddharth Ahuja, which works with any MCP-capable client — Claude Desktop, Claude Code, Cursor and VS Code. It has two halves: a Blender add-on that opens a socket server inside Blender, and an MCP server that the client talks to. You need Blender 3.0+, Python 3.10+, and the uv package manager (installed via uv's own installer — not pip).

1. 01Install uv: brew install uv on macOS, curl -LsSf https://astral.sh/uv/install.sh | sh on Linux, or the PowerShell installer on Windows (then add it to PATH).
2. 02Download addon.py from the blender-mcp repository.
3. 03In Blender, open Edit → Preferences → Add-ons → Install…, select addon.py, and enable “Interface: Blender MCP”.
4. 04Add the MCP server to Claude Desktop under Settings → Developer → Edit Config, using the JSON below.
5. 05Restart Claude, then in Blender press N to open the 3D View sidebar, switch to the BlenderMCP tab and click “Connect to Claude”.
6. 06Ask Claude something small first — “what objects are in the scene?” — to confirm the two-way connection before you start building.

{
  "mcpServers": {
    "blender": {
      "command": "uvx",
      "args": ["blender-mcp"]
    }
  }
}

If Claude reports that it can't find the command, replace "uvx" with its absolute path (for example /opt/homebrew/bin/uvx on macOS). The add-on communicates over localhost:9876 by default; BLENDER_HOST and BLENDER_PORT override that if you're running Blender on another machine.

The same MCP server works with any MCP-capable client, so you are not tied to the Claude desktop app. Claude Code, Cursor and VS Code all accept the identical mcpServers block — Cursor takes it under Settings → MCP. On Windows, some clients need the command wrapped as "cmd" with args ["/c", "uvx", "blender-mcp"].

Practical difference: the desktop app is the smoothest for conversational modelling, while Claude Code is better when the Blender work is part of a larger project — generating assets from a spec file, batch-processing a folder of .blend scenes, or scripting a pipeline where Blender is one step among several.

Through blender-mcp, Claude gets a real toolset rather than a text box. In practice that means:

• –Inspect the scene — list objects, read their properties, and take viewport screenshots to see the result of its own work
• –Create, delete and modify objects, including modifiers and transforms
• –Create or apply materials and colours
• –Execute arbitrary Python (bpy) code inside Blender — the escape hatch that makes everything else possible
• –Pull in assets: download models from Sketchfab and Poly Haven
• –Call out to mesh generators such as Hyper3D Rodin and Hunyuan3D for organic shapes it can't build procedurally

That last pair matters more than it looks. Claude builds geometry as code, so a chest of drawers or a spiral staircase is comfortably within reach, while a gnarled tree stump is not. Delegating the organic pieces to a generator and assembling the scene itself is the workflow that plays to its strengths.

Because arbitrary Python execution is on the table, save your file before a long session. A wrong bpy call can wipe a scene, and Claude has no undo you can trust.

This is the question the arena exists to answer, and we deliberately don't guess at it. Season 1's Round 1 put four Claude tiers — Haiku 4.5, Sonnet 5, Opus 4.8 and Fable 5 — through the same 12 benchmark prompts, each generating its entries independently, one shot, no retries and no cleanup. Their outputs face each other in blind votes on this site, and the standings move only when real ballots arrive.

What we can say without vote data: the cheaper tiers are noticeably faster and cost a fraction as much, which matters because agentic Blender sessions burn through usage quickly; the larger tiers hold multi-step scene state better over long sessions. Whether that translates into better-looking geometry on a single prompt is exactly what the blind votes are testing.

• –Claude can't see Blender — check the BlenderMCP sidebar tab actually says it's connected; the add-on's socket server has to be started from inside Blender, it doesn't autostart.
• –“Command not found: uvx” — use the absolute path to uvx in the config; GUI apps don't inherit your shell PATH.
• –The first command fails, later ones work — a known quirk of the connection handshake. Retry once before debugging anything else.
• –Long builds drift — break the task into steps and ask Claude to screenshot the viewport between them, so it corrects against what's actually there instead of what it assumed.
• –Usage burns fast — scene sessions are tool-call heavy. Start on a cheaper tier for blocking out, escalate only for the parts that need it.