Blender Studio MCP

Lets Claude and other AI models make good-looking 3D art in Blender. Project page: https://just-rice.github.io/blender-studio/

Blender can already be scripted, but an AI writing raw scripts tends to produce flat, grey "programmer art". Blender Studio gives the AI what a photographer has: ready-made lighting rigs, real materials, a camera that frames the subject by itself, and above all the ability to render, look at the picture, and fix it.

Each picture was built and photographed using only the tools below; see examples/showcase.py.

Straight out of the test suite (tests/motion.py), unretouched. A sample video is in docs/walk-and-jump.mp4.

• Works on macOS, Windows and Linux.
• Works with any AI app that supports MCP (Claude Code, Claude Desktop, Cursor, and others).
• Free. No account and no API key. Blender runs on your own computer.
• Blender starts by itself in the background; there is no add-on to install.

Install

1. Install Blender 4.2 or newer.
2. Install uv.
3. Add the server to your AI app.

Claude Code

claude mcp add blender-studio -- uvx --from git+https://github.com/Just-Rice/blender-studio blender-studio-mcp

Claude Desktop, Cursor and other apps (add to the app's MCP settings file)

{
  "mcpServers": {
    "blender-studio": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/Just-Rice/blender-studio", "blender-studio-mcp"]
    }
  }
}

Then ask for something: "Make a product shot of a ceramic mug on a warm backdrop."

What the AI can do

Tool	What it does
create_object, transform_object, delete_objects	Build with basic shapes
run_python	Anything else, using Blender's own scripting
apply_material	24 realistic surfaces: ceramic, gold, glass, wood, fabric, neon and more
create_rock, create_terrain, create_blob, create_limbs	Natural and organic shapes: stones, hills, clouds, creatures, trees
sculpt, boolean	Roughen, smooth, fuse and cut existing shapes
create_character	Stylised people in three styles (toy, low-poly, blocky) with working joints
create_human	Realistic adult bodies via the free MPFB add-on, installed on first use
pose_character, animate_character	11 poses (wave, sit, run...) and 6 motion clips (walk, jump, dance...)
list_animations, play_animation	Use the motion that comes with imported characters (Mixamo, Ready Player Me, any FBX/GLB)
animate	Turntables, camera orbits and dollies, bounce, spin, float, pop-in, light fades
drop_objects, drape_cloth	Real physics: things fall and settle, fabric folds over objects
render_animation	MP4 video, with a strip of frames returned so the AI can check the motion
add_polish	Soft edges and smooth shading, so objects catch the light
add_backdrop	Seamless photo-studio sweep or a floor
studio_lighting	Seven complete rigs, sized and aimed at the subject automatically
frame_camera	Nine shot types with automatic framing, lens choice and background blur
set_look	Seven overall moods: clean, cinematic, pastel, moody, vivid, flat, monochrome

Pictures and scene files are saved in a BlenderStudio folder in your home folder.

Try it with another AI

examples/gemini_demo.py lets Google Gemini drive a cut-down set of the tools, to show nothing here is tied to Claude. It fits inside Gemini's free daily allowance.

GEMINI_API_KEY=... uv run python examples/gemini_demo.py out "a toy rocket on a launch pad"

Does it help?

A benchmark gives the same request to the same model with plain Blender scripting and with these tools, and a blind judge compares the pictures. In the latest round, run after fixing what the first round exposed, the Blender Studio picture was judged better in 9 of 12 pairs, with an average of 6.1 against 4.7 out of 10. Every Claude model tested averaged higher with the tools, and no tool call failed, though the tools cost more tokens to read. The pictures are on the project page and the full numbers, limits and the first round are in bench/RESULTS.md.

Options

Set these as environment variables in the MCP settings if you need them.

Variable	Purpose
BLENDER_PATH	Where Blender is, if it is not found automatically
BLENDER_STUDIO_WINDOW=1	Show the Blender window and watch the scene being built live
BLENDER_STUDIO_HOME	Where to save renders and scene files

How it works

The MCP server starts Blender with a small bridge script loaded, then sends it one command at a time over a local connection that only your own computer can reach. Nothing is sent anywhere except asset downloads from Poly Haven when the AI asks for them.

Safety

run_python lets the AI run code inside Blender, with the same access to your files as any program you run. Use it with AI apps you trust, and keep their approval prompts switched on.

Development

git clone https://github.com/Just-Rice/blender-studio
cd blender-studio
uv run python tests/smoke.py              # builds and renders a test scene
uv run python tests/motion.py out         # sculpting, people, animation, physics, video
uv run python tests/mcp_client.py         # talks to the server the way an AI app does
uv run python examples/showcase.py out    # renders the three pictures above
uv run blender-studio-mcp                 # runs the server

Claude Code users can also copy skill/blender-studio into ~/.claude/skills/ for a short workflow reminder.

Limits

• No brush sculpting: organic shapes come from noise, blobs, skins and cuts.
• create_human bodies wear a painted-on outfit and have simple faces. Real clothing and hair need MPFB's separately downloaded asset packs, which this does not manage.
• Motion clips are simple procedural loops, not motion capture. For lifelike acting, import a character that already has animation.
• No fluid simulation.
• Tested on macOS with Blender 5.2 only.

MIT licence.