Blender MCP lets an LLM control Blender through the MCP protocol, using Blender's scripting support to create scenes from prompts.
2
Keeping an MCP's tool set small and distinct helps models choose the right action, while adding every possible feature can make the user experience worse.
3
MCP clients could eventually coordinate Blender, Unity, Ableton, asset APIs, and other creative tools from a user's intent rather than requiring them to learn each application's interface.
Summary
Siddharth Ahuja explains how he built Blender MCP to let clients such as Claude or Cursor control Blender through the Model Context Protocol. Blender's scripting support lets an LLM write and execute code for modeling, asset retrieval, scene setup, animation, and other tasks. Ahuja describes demos involving dragons, generated assets, terrain, games, racing scenes, and reference-image-based rooms. He is direct about the limits: current models have a poor understanding of 3D, and a larger collection of tools can confuse the model. His broader argument is that MCP clients could become an orchestration layer across creative software. A user might ask to make a game, while the client calls Blender for assets, Unity for gameplay, APIs for resources, and Ableton for music. This could let creators work from their intended outcome without first learning every underlying tool, although the quality of the generated work still has room to improve.
Blender's complexity creates a strong case for natural-language control
Ahuja describes Blender as a generalist 3D tool that can import assets, animate them, create art, and export work to a game engine. Its interface has many tabs and settings, which made the tool difficult for him to learn as a designer and engineer. He uses the classic beginner donut exercise to show the barrier: completing it takes about five hours. When he saw that MCP could let an LLM talk to tools, he wondered whether it could make Blender easier to use. The goal became letting someone describe a scene instead of manually working through Blender's interface.
Blender MCP turns a prompt into scripts and tool calls inside Blender
Blender MCP connects a client such as Claude or ChatGPT to Blender through the MCP protocol. MCP lets Blender expose the actions it can perform, so the client can understand and call those tools. A Blender add-on then executes the scripts that the LLM sends. In Ahuja's example, a prompt to make a dragon guarding a pot of gold produces an isometric room, a dragon with wings, and the gold in about five minutes. The result is imperfect, but it is faster than building the scene manually. Blender's ability to run scripts is what lets the model control modeling and scene creation.
External asset services extend what a Blender prompt can produce
The system can connect the LLM to asset sources such as Rodin, Sketchfab, and Poly Haven. A prompt can request an asset, such as a zombie, and the client can obtain or generate it and place it inside Blender. Ahuja says the client handles much of the coordination, while Blender's support for downloading and importing assets makes the workflow easier. Because clients can connect to any API, the same pattern is not limited to one asset provider. The user can describe what is needed, and the system can retrieve or create the asset as part of the scene-building process.
A small, distinct tool set makes MCP more reliable
Ahuja had to refactor Blender MCP after giving it about 14 or 15 tools. The model became confused about which action to choose when tools included operations such as creating a cube, creating a sphere, executing code, and retrieving an asset from Rodin. Since tool selection is somewhat nondeterministic, he found that each tool should be clearly different and that the overall experience should stay lean. His advice is not to add features simply because they are possible. Blender MCP works partly because a generalist tool can do a lot without exposing an unnecessarily large set of choices.
Better underlying models quickly improve the same MCP
Ahuja says current language models still have a poor understanding of 3D, and 3D models are not yet great. He also saw rapid improvement while Blender MCP was being released. In his example, the arrival of Gemini 2.5 made the system about three times better. This changes how he thinks about building the next MCP. The tool design has to account for current model weaknesses, but improvements in the underlying models can change what the same interface can accomplish without requiring a complete redesign.
Prompting lowers the time needed for scenes, animation, and games
Ahuja shows examples of people using Blender MCP to make scenes in minutes. One scene uses AI-generated assets for a magical mushroom setting, while another creates and animates a cat. A reference image of a living room is used to recreate the room with appropriate assets. Another example generates terrain and complex node-based textures from an image. Someone else created a game set inside lungs, with assets, glossy materials, and camera animation. A racing track example combines Blender MCP with Runway to turn an animated scene into a video clip. He also compares the beginner donut workflow, which took five hours, with a one-minute prompt-based version.
The client could coordinate an entire creative workflow
Ahuja extends the idea of the client as an orchestrator. Someone who wants to make a game usually wants the game, rather than an experience learning Unity. The client could call Blender to create assets, Unity to add collisions and game logic, external APIs to retrieve resources, and Ableton to create the soundtrack. He demonstrates this direction by combining Blender MCP with an Ableton MCP. The prompt creates a dragon, gives it sinister lighting, and asks Ableton to make music for it. The quality is limited for now, but the pieces can already be connected through one interface.
Creative work may shift toward directing systems through intent
Ahuja asks whether people will increasingly interact with an LLM instead of learning the complex interfaces underneath creative tools. He suggests that creators may become more like orchestra conductors, where communicating the vision and coordinating the parts matters more than knowing how to operate every instrument directly. He points to MCPs for tools including PostGIS, Houdini, Unity, and Unreal Engine as signs that this pattern is spreading beyond Blender. His claim is that more people could create when the software responds to what they want to make, although he does not present the current output quality as finished.
"I feel like MCPs are going to be a fundamental glue that holds this thing together."12:48
Who should watch
You are building an MCP server and need practical guidance on keeping its tools distinct and its interface small.
You work with Blender or other creative software and want to see how prompts can handle assets, scenes, animation, and simple games.
You are thinking about cross-tool creative workflows and want an example of a client coordinating 3D, game development, music, and external APIs.
