最佳 Claude + Blender 配置：MCP 安装与 MCP Skills

The required setup

You need two pieces. Add the optional skills only when a use case asks for them.

Prerequisites

• Blender 5.1+ — the Claude UI says 4.2+, that's wrong; the MCP add-on requires 5.1.
• Claude Desktop (Mac or Windows). The browser app at claude.ai does not support connectors.

1. Official Blender Connector

Blender side. Open the Blender Lab MCP page next to Blender. Drag the install link into the Blender window — Blender prompts to allow the lab extension repository, accept. Drag the same link in a second time and the MCP add-on installs. Confirm via Edit → Preferences → Add-ons → search “MCP”.

In the 3D Viewport, press N to open the sidebar, find the BlenderMCP tab, click Connect to Claude, and tick Auto-start. The server runs at localhost:9876.

Claude Desktop side. Settings → Connectors → Browse → search Blender → Install. Open the connector's Configure page; you'll see 19 read tools and 7 write/delete tools. Set the common read tools to Always allow so permission popups don't break the flow.

Smoke test. In a fresh chat: “Are you connected to Blender? Delete the default cube.” The cube should vanish.

Anthropic's tutorial for the connector is here: claude.com tutorial.

2. fal.ai MCP

fal.ai is the cloud half of the stack — image, video, 3D, and audio generation across 1000+ models. PATINA (PBR textures), Flux, Hunyuan3D, Kling, Wan, and so on. You'll need a fal.ai key (free credits to start; production usage runs on metered billing).

If you use Claude Code, one command:

claude mcp add --transport http fal-ai https://mcp.fal.ai/mcp \
  --header "Authorization: Bearer $FAL_KEY"

For Claude Desktop, open Settings → Developer → Edit Config and add this to claude_desktop_config.json:

{
  "mcpServers": {
    "fal-ai": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://mcp.fal.ai/mcp",
        "--header",
        "Authorization: Bearer ${FAL_KEY}"
      ],
      "env": { "FAL_KEY": "<your_fal_key>" }
    }
  }
}

Save, then quit Claude Desktop fully from the menu (not just the close button) and relaunch.

Docs: fal.ai MCP setup.

Why fal.ai is my main recommendation

You can ship complete videos with just the two pieces above — that's the whole point of recommending them as the baseline. Most third-party 3D AI services have their own dedicated Skills and MCPs, and they're useful, but if you're picking one to start with, fal.ai gives you the broadest coverage in a single connector: PBR materials, image and video generation, Hunyuan3D and Tripo backends, all from the same chat session, all priced per call.

Optional add-ons

Add these only if you specifically want them. Each one extends the stack with one more provider or one more set of tools.

Custom Skills follow the same shape

Most third-party Claude Skills install identically: a folder containing SKILL.md plus a .env with your API key, dropped into ~/.claude/skills/. The 3D AI Studio skill is a worked example.

Claude Code:

git clone https://github.com/t4ai-workshops/3d-ai-studio-api-skill \
  ~/.claude/skills/3d-ai-studio-api/
echo "3D_AI_STUDIO_API_KEY=<your_key>" \
  > ~/.claude/skills/3d-ai-studio-api/.env

Claude Desktop: download the repo as a ZIP, put 3D_AI_STUDIO_API_KEY=<your_key> in a .env next to SKILL.md, re-zip the folder so SKILL.md is at the archive root, then Customize → Skills → + → Upload a skill.

Same procedure for any Skill that ships as a SKILL.md plus environment variables.

Verify the stack

In a fresh Claude chat, ask:

Expected: blender, fal-ai, plus whichever optional skills you installed. If something's missing, the most likely culprit is a config typo or a half-restarted Claude Desktop — quit fully and reopen.

Examples I tested

Seven prompts I've personally run end-to-end. Copy-paste, swap object names where needed, ship.

1. Bake high-poly to low-poly

The classic game-asset workflow, fully automated. Claude finds the meshes, sets up Selected to Active, hides everything else from render so it doesn't pollute the AO, bakes Normal and AO at 2K, saves the PNGs, and wires them into the low-poly's material — Normal through a Normal Map node into BSDF Normal, AO multiplied into Base Color.

2. Generate a 3D character from images

Image-to-3D plus auto-import and assembly inside the live scene. Three reference images go in (pig head, body in a suit, top hat), three GLBs come out via the 3D AI Studio Skill, and Claude stacks them into a character without resetting your scene.

3. PBR materials with fal.ai PATINA

My favourite demo. Four tileable PBR materials generated in parallel and wired up to a small corner scene: herringbone parquet for the floor, William Morris-style botanical wallpaper inside, London red brick on the wall's exterior, walnut for the table.

4. Toon shader with outlines

Multi-mesh character to cel-shading on top of the existing diffuse maps, plus a Solidify-based outline. Includes a guard against the classic “outline explodes on hair and bows” failure mode.

5. Procedural iridescent paint shader

Replace a body material with a procedural rainbow-iridescent shader driven by Layer Weight (Facing) and a Color Ramp — the kind of pearl/chameleon paint you see on tuned cars or beetle shells.

6. Multi-stage object animation

Auto-split a merged keys mesh into per-key islands, then animate the keyboard assembly. The body drops in first, then keys fall from above sweeping left-to-right across all rows simultaneously, with row clusters identified automatically by depth.

7. Geometry Nodes: scatter plants on a terrain

Claude builds a real Geometry Nodes graph end-to-end — bounding box, grid of points, jitter, raycast onto the terrain, instance random plants from a collection. Note the explicit handling for Blender 4.x vs 5.x raycast API differences in the prompt.

Where to start

If you're installing this fresh, run example 3 (PATINA materials) as your first real test. It's visually striking and exercises fal.ai, the Blender connector, and material wiring all in one prompt — if that works, the rest works.