AI Engineer World's Fair 2025
Blender MCP and the Future of Creative Tools - Siddharth Ahuja
Read the talk
Blender MCP: Turning Creative Intent into Tool Operations
Blender’s scripting interface lets a language model turn a scene description into operations. Siddharth Ahuja’s experiments show why distinct tools matter, and how the same approach can connect generated objects, animation, and music.
From a talk by Siddharth Ahuja
Why a donut becomes an interface problem
Making a donut sounds like a modest creative goal. In Blender, getting there can first mean learning a substantial interface. Siddharth Ahuja came to the problem through experimentation after running a product studio: he liked what the application could do, but had never mastered its controls.
Blender is a general-purpose 3D application. It can import assets, animate them, create artwork, and export work to a game engine. Those capabilities come with tabs across the top, properties on the right, and more options nested inside them. Ahuja jokes that navigating all those controls had kept him from becoming a good 3D artist.
When he encountered Model Context Protocol, or MCP, its promise of connecting a language model to tools suggested a different entry point. Could the model handle the operations while he described what he wanted?
Ahuja describes a beginner donut course he followed as a five-hour exercise. His frustration was the distance between a simple intention and the knowledge needed to carry it out. The question behind Blender MCP was whether prompting could shorten that distance.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A dragon that gets the brief
Blender MCP connects a model application, such as Claude or ChatGPT, to Blender so that a prompt can produce scene operations. Ahuja’s opening demonstration asks for a dragon guarding a pot of gold. The user supplies the scene’s intent; the model application works out how to construct it.
Ahuja laughs at the result: it understands the assignment, even if it is hardly exceptional artwork. There is an isometric room, a dragon with wings, and a pot of gold. Ahuja estimates that the dragon scene took about five minutes to make. He thinks the same attempt would have taken him much longer manually.
Getting a recognizable result and producing polished artwork are separate achievements. This dragon demonstrates the first: a description has become an arrangement of objects inside a complex creative application.
At the time of the talk, Ahuja reported approximately 11,500 GitHub stars for Blender MCP. He also reported more than 160,000 downloads. Other people were already using it and building on it, expanding the experiments beyond what he could make himself.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The model selects tools; Blender executes the work
The connection is simpler than the resulting scenes might suggest. A model application communicates with Blender through MCP, which provides a standardized way to expose available capabilities. Ahuja’s sketch puts the application between Blender and external asset services: it can coordinate work inside the local scene and requests to outside APIs.
Tool descriptions tell the model which operations are available. The model selects an operation, and the application invokes it; MCP itself does not choose what to make.
On the Blender side, Ahuja’s add-on executes scripts supplied through the bridge. That is the crucial connection between language and geometry: generated code becomes operations inside Blender. The model application interprets the request, while Blender executes the scene work.
Assets extend the available vocabulary beyond objects constructed directly through scripts. Rodin supplies AI-generated assets, while Sketchfab and Poly Haven provide other asset sources. The model application can coordinate obtaining an object and bringing it into the scene.
Blender makes this practical because it already supports scripting and can download and import assets. The integration can use those existing capabilities instead of reproducing each action as a sequence of clicks through the interface.
For example, a user can describe a zombie and ask for a generated asset that fits the description. The client coordinates that request with the external service and the work inside Blender. This is its role as an orchestrator: the creative request may require several capabilities, but they can contribute to one result.
Ahuja presents the separation between application, model, and creative tool as a source of flexibility. He names Claude, Cursor, Windsurf, and LangChain as possible environments, and points to choices across Claude, ChatGPT, and Gemini for the model side. His design ambition is to let people bring their preferred environment and model to the same creative workflow.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Distinct tools reduce selection ambiguity
The first implementation lesson is how much work an existing scripting interface removes. Claude’s ability to write code becomes useful for modeling and asset handling because Blender can execute that code. A general execution capability can reach many parts of the application without requiring a separate integration for every manual operation.
But exposing more operations does not automatically make the model more effective. Ahuja reports repeatedly refactoring an integration that had roughly 14 or 15 tools because tool selection became confused. The problem was deciding which capability to use for a request.
His examples show why the choices can overlap:
| Exposed capability | Route to a result |
|---|---|
| Create a cube or sphere | Request a primitive shape directly. |
| Execute code | Use a script to construct scene content, including primitive shapes. |
| Retrieve an asset from Rodin | Obtain a generated asset rather than construct it directly. |
A request for a simple shape can fit both a dedicated creation tool and general code execution. The model has to choose between routes that may lead to the same visible result.
Ahuja’s response is to keep the interface lean and make each capability clearly distinguishable. Distinct responsibilities make tool selection easier. The useful lesson is to reduce ambiguity between operations, rather than adopt his tool count as a universal limit.
That principle carries into the product experience. An open-source project can attract many feature requests, but adding every possible feature can erode the lean, general-purpose workflow that made it useful.
Meanwhile, the underlying models improve. Ahuja still finds their understanding of 3D limited, but points to Gemini 2.5 as an example of progress he noticed during the project’s development. A broadly useful scripting bridge can benefit from that progress without needing a new dedicated feature for every kind of scene.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Generated assets, animation, and image references
Once other people started using the bridge, its possibilities became more concrete. Ahuja begins the showcase with his own scene assembled from newly generated assets, including a magical mushroom. He describes the objects he wants, tells Claude where to place them, and captures the resulting composition.
Ahuja reports making the magical-mushroom scene in two minutes. He emphasizes that the assets were generated for the scene: prompting covers both obtaining the objects and arranging them, rather than only positioning a collection he already had.
A separate cat example adds another operation. Its creator obtains generated assets and then animates them a little, giving the scene a sense of life. Ahuja says the cat workflow can be completed in less than an hour. The result is no longer just a composition to photograph; its objects can move.
The next example changes the input. A creator supplies an image of a living room, and the model uses that reference to guide reconstruction and obtain appropriate assets. Ahuja describes the living-room reconstruction as taking minutes.
A reference image gives the workflow a visual target. Instead of describing every object and spatial relationship in words, the creator can show the arrangement they want the scene to resemble.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Terrain nodes, game scenes, and animated materials
Another creator works in Cursor with an image reference to generate terrain in Blender. This reaches into a more specialized part of the application: nodes that configure textures and surface bump or normal detail.
Those node arrangements have their own learning curve. Because Blender can execute code, the model can configure them as part of creating the requested terrain. The scripting bridge reaches beyond placing objects into the machinery that determines how their surfaces look.
Ahuja is particularly delighted by the next game’s unusual premise: the player enters someone’s lungs, collects bone fragments, and reconstructs a rib cage to win. Blender MCP was used to set up the scene and create assets. Its contribution here is the environment and objects that support the game’s peculiar objective.
A further showcase focuses on a glossy, iridescent material and an animated camera. Ahuja says this material and camera animation were produced through prompting in minutes. The request now reaches beyond an object’s shape into its surface appearance and the camera movement used to present it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A Blender racing scene becomes a Runway clip
The racing example begins with a prompt to create a track and animate cars. The Blender scene contains a circular racetrack, cars, colored obstacles, trees, and stands. The creator then places the camera so the racing action starts to feel like a movie shot. This is a constructed 3D scene whose objects and framing can be directed before the next production step.
The creator also animates the camera, then uses Runway to transform the Blender-derived material into the final displayed clip. Blender establishes the scene and motion; Runway produces the transformed footage.
That division opens the workflow to filmmakers as well as 3D artists. A controllable scene can become an intermediate step in making a shot, even when the Blender render is not the final visual medium. The track, cars, and camera movement supply structure for the subsequent video-generation work.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Returning to the donut
Ahuja returns to the opening exercise with donuts made from a short prompt. Ahuja estimates about a minute for the prompted donut example. He also acknowledges that the result could be much better.
The contrast with his five-hour beginner course concerns the route to a starting result. Following a tutorial teaches operations; prompting produces something to inspect and refine. They do not establish equal quality or equal skill acquisition, but the prompted example changes what a newcomer can attempt before mastering the interface. A simple idea can now become a visible object that gives the next prompt something concrete to work on.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A workflow organized around creator intent
Ahuja then extends orchestration beyond Blender. If a model application can coordinate a local creative tool and external asset services, it could also coordinate several creative applications.
The organizing principle is creator intent. People come to Unity because they want to make a game, and to Ableton because they want to make music. Learning the software is a means to that outcome. The challenge is to capture what someone wants to create and translate it into work across applications while still matching their vision. Ahuja sees MCP connecting those tools, with the language model interpreting the request and coordinating the work.
In his proposed game workflow, the user describes the game to the model application. The application would coordinate several responsibilities:
- Blender: Create the assets needed for the game.
- Unity: Assemble the game, add collisions, and implement its logic.
- External services: Supply additional assets or animation.
- Ableton: Create music for the characters and game.
This is an envisioned production chain. The creator would direct one project while the application chooses which underlying tool should perform each part.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A dragon and its soundtrack
To demonstrate a smaller version of that idea, Ahuja combines Blender MCP with his companion project, Ableton MCP. The request is to create a dragon and give it a soundtrack. Ableton MCP supplies the connection through which the model application can request music creation, alongside its existing connection to Blender.
The result has two coordinated parts. Blender contains the described dragon, with sinister lighting appropriate to a villain. Ableton creates the soundtrack. The character brief influences both the visual scene and the music, while each application produces its own output.
Ahuja is candid that the quality is still limited. What excites him is the ability to stitch the pieces together: one creative request can guide work that would otherwise require separate operations in two very different interfaces.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
What creators still need to direct
These experiments leave Ahuja with a question about interfaces: will more people work primarily through a model application, without learning every complex tool beneath it? He sees new creators entering that way already.
His analogy is an orchestra conductor. The creator needs to articulate the intended result and coordinate its parts, even when they do not personally operate every instrument. Expressing a vision clearly becomes central to making it executable: which pieces belong together, what each should contribute, and how the result should feel.
Blender MCP and Ableton MCP are his concrete examples, but Ahuja also reports seeing integrations for PostGIS, Houdini, Unity, and Unreal Engine. He expects that expanding set of connections to let more people create across media. His enthusiasm rests on the possibility that someone with a clear idea could direct the necessary tools before becoming an expert in each one.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
The speaker's bridge from an LLM client to Blender's scripting and scene tools.
The second creative tool used to add a soundtrack to the closing dragon demonstration.
Further reading
Current host/client/server vocabulary; protocol examples in this version postdate the talk.
Read the complete timestamped transcript
- 0:01
Hey everyone, this is Siddharth. Welcome to my talk about the Blender MCP. Uh, for those of you who don't know what Blender is or MCP is, I'm going to go through it in a bit.
- 0:09
But I'm gonna talk about how my experience was building this, what I learned, and where I see it going forward, uh, especially for more creative tools even outside of Blender.
- 0:20
So maybe a little bit about me before we start. I've been a designer and an engineer for about eight years now. I just like tinkering with stuff. I ran my own product studio for a while, and this Blender MCP was a result of some experimentation.
- 0:35
Um, just... but just before we jump into it, you know, a lot of you might be asking: What is Blender? So Blender is just a 3D tool. Think of it as a generalist 3D tool.
- 0:44
You can import assets, you can animate them, you can export them out to like a game engine, so you can do all sorts of stuff in Blender. You can create art, right?
- 0:54
And the UI kind of looks like this. It's pretty complex if you see all the different tabs everywhere. Uh, on the top over here, there's a bunch of things.
- 1:03
On the right over here, there's a bunch of things. Each of these has like 10 tabs inside it. So it gets pretty complex, which is the reason I myself could never be a good 3D artist. [chuckles]
- 1:14
I was too lazy to go through all of these different tabs. And, uh, uh, I, I guess that was kind of my motivation when I saw MCP and, uh, I saw that MCP could let your LLM talk to any tool.
- 1:28
My first response was, "Oh, wait, will this make Blender easier?" You know, will this make a historically complex tool easy to use? And so just to give you an idea of how it used to be or even still is, it takes five hours to build this donut, and this is a very classic beginner's course that even I
- 1:47
did. But, uh, th- this is the status quo, right? It takes five hours to go through this and build a tutorial, and it should not be this hard, right?
- 1:57
And so the whole idea behind Blender MCP, uh, is, is essentially to let your LLM, like Claude or ChatGPT or whatever, right? Let them talk to Blender, which is a complex tool, and let the LLM control Blender so that it can make 3D scenes for you, and you just have to prompt.
- 2:15
So in this example, this is a demo I posted. Essentially, I'm just saying, "Make a dragon," right? "Have it guard a pot of gold," and let's see what it does.
- 2:44
Right. So you can see it kind of gets the brief, [chuckles] but also not the best. But it, it made an isometric room. It has a dragon. It kind of gave it wings.
- 2:53
There's a pot of gold. It, it made it in what? I think about five minutes. I would have taken a lot longer to do something like this, to be honest.
- 3:03
Uh, so this is, this is where I think I'm gonna talk about a little bit more. So Blender MCP launched to about 11 and a half stars on GitHub at this moment.
- 3:13
Uh, and it has almost, I think, way more than 160k plus downloads. A lot of people have used it. A lot of people have built on top of it, and I'll show some examples in a bit.
- 3:25
But how, how does it work exactly, right? So it's actually way more straightforward [chuckles] than it seems.
- 3:32
So it's just essentially your client, which is Claude or Cursor or whatever, it connects to Blender through the MCP protocol, right? If you haven't seen what an MCP is, I feel like a lot of people are...
- 3:45
have been talking about it even in this conference itself. But it helps you. It's just a standardized protocol that allows Blender to show to the client, "Oh, this is what I can do."
- 3:56
Like, so that's the tools part of the MCP that, "Oh, these are the tools I have, and this is what I can do." And so the LLM understands that and then uses those tools.
- 4:04
So how it uses those tools is in Blender, there's an add-on that I've made, uh, to execute the scripts that the LLM tells it to. So for example, Claude will say, "Make a dragon," and it will call the exact right tools to make those...
- 4:19
make that dragon in Blender. Another big part of it is the assets, right? So the industry standard like Rodin, which is an AI-generated asset, Sketchfab and Poly Haven, things like that are now connected to the LLM, and that can basically now seamlessly, if you write a prompt to generate an asset inside Blender, it can just do that,
- 4:39
right? So the client does all the heavy lifting, essentially. Yeah. So I think Blender is a pretty big part. I think this reason this is possible in the first place is because Blender has scripting, right?
- 4:50
So code can be executed inside Blender, which is a big part. Blender is pretty flexible with also downloading assets and importing, and so that makes the life like much easier, honestly.
- 5:02
With assets, yeah, I think the cool thing here is that this is just the start, but clients can connect to any API. So for example, right now, you can write any prompt you want, and it will get the right asset for you.
- 5:14
You can say, "I want a zombie," and it'll get the right asset for you, AI-generated on the spot. You can describe it however you want. And then the client is kind of the orchestrator, right?
- 5:25
So the cool thing about this is I think that you can use it in any tool you want. You know, it could be Claude, it could be Cursor, it could be Windsurf, LangChain, whatever you like.
- 5:35
Uh, and you can use any language you want with it, right? You can use, uh, uh, Claude and ChatGPT and Gemini and whatever you like. So I think this is how it broadly functions together.
- 5:47
If you want more detail, there's al- always the GitHub. But, uh, I think some learnings while I was making this was that a tool with scripting can remove a lot of the heavy lifting.
- 5:58
Like I mentioned, the Blender thing, right? Because Claude is good at writing code, it can very easily translate it to Blender, and Blender can execute that code. And that can be for modeling, for getting the right assets, anything.
- 6:12
Um, the second part I learned was that MCPs get really confused with tools. So, uh, I learned this the hard way where, where I had to refactor the Blender MCP quite a few times because I had about, I think, 14 or 15 tools, and it kept getting confused on what to use.
- 6:28
So tools can be something like create a cube, create a sphere, execute code, uh, get this asset from Rodin, for example, right? So how does it know which one to choose?
- 6:39
Because it's a little bit non-deterministic, right? So in that case, it's always better to kind of keep the end user experience in mind and making sure it's really lean and every single tool is different from each other so that Claude knows exactly which tool to pick.
- 6:56
Um, and a good segue from that is don't bloat the UX just because you can. I think this is more a learning from making it open source, which is just be-- I think there's a lot of features that can be added, but I think the reason Blender MCP works is because lean and a bit of a generalist
- 7:13
tool, so it can get a lot done. And another learning has been that the models do get better, the underlying models. So while 3D models aren't great yet, and LLMs have a very poor understanding of 3D in general, they're getting way better.
- 7:28
So even in the time I released it to the time it was out, Gemini 2.5 came out, and that made it already 3X better, right? So I think those are some things to keep in mind when I'm building at least the next MCP.
- 7:44
Now, sure, now I've put this out in the wild, and now comes the creative tools part, right? What's, what's changed? Uh, how are people using it? So historically, like I said, it's been pretty complex to use 3D tools, and now suddenly the barrier of access has reduced.
- 8:04
And so this is a scene I made in two minutes [chuckles], you know, just got the right a-assets. AI generated all of them. These assets don't exist otherwise. They're all AI generated and just place them, tell Claude how to place it, tell them, "I need a magical mushroom," and it just creates everything from scratch, and I can
- 8:20
just take a picture of it, essentially. This is, uh [upbeat music] another one.
- 8:29
Where... This is another one where they're creating a cat, and they're also getting all the right AI generated assets, and then they animated it a little bit to make it feel real.
- 8:39
And you can do all this in less than an hour now, where this would have taken way more time earlier, right? This is another one [upbeat music]
- 8:48
where, where they're providing reference images, right? So they just gave an image of a living room, and now it's recreating that, getting all the right assets. Again, this would have taken quite a long time earlier to just make this, but now it's a matter of minutes.
- 9:04
Some other cool things. This person, uh, on Twitter, they're generating terrain. So they have an image which they gave to the Blender MCP, uh, and inside Cursor, they're trying to create, create this game on Blender.
- 9:16
And so it generates the right terrain for you. It even uses these nodes, right? Which is, which it itself has a learning curve to set up these really complex textures and normal bumps and so on.
- 9:29
And so it does that, all of that for you just because Blender can execute code, right?
- 9:34
This is... I love this one. This is [upbeat music] a game that someone created. This is essentially...
- 9:41
The idea is that they go into someone's lungs, and they have to collect bone fragments, um [chuckles], to, to win the game, to connect the rib cage together. And so Blender MCP, again, was used to set the scene to create assets.
- 9:53
I'll, I'll just play it shortly. [upbeat music] Yeah.
- 10:12
I'll just leave it at that. Um, so you can create games now [chuckles]. And this is, uh, completely done in Blender MCP. It's a pretty high fidelity of, you know, glossy, this iridescent material and a camera animation all done in minutes by just prompting through AI, right?
- 10:30
Uh, so these are the things that are being unlocked. Um, this is, uh, another cool thing that I saw that people have been using it for. So this person prompted the MCP to create a racing track and then animate the cars.
- 10:43
So you can see on the side the MCP is running and, um, on the right side, this Blender is running essentially. And so this person created this, and now you can see if I skip to the end a little bit, this person now placed the camera angle such that the cars feel like they're [chuckles], feel like a
- 11:04
movie, and then they used Runway to convert them to an actual clip. So here, here you can see it. Whoops.
- 11:13
Yeah, there you go. Yeah. So they animated the camera, and now this is through Runway. So it's another medium that's been unlocked, basically, where you're... Even filmmakers or creators, all kinds of creators can use this, not just 3D artists.
- 11:27
And this is ju- [chuckles] just an example of some donuts, which are one short prompt. This is just, you know, like in a minute. Uh, it can be ma- maybe, uh, made way better, of course.
- 11:39
But, uh, you know, as a start from five hours to a minute is not too bad. So this is essentially unlocking a whole new
- 11:49
world of creators to just come in, and when the barrier has been so high, now it's just gone, and you can just write what you want, what's in your head, and get there with just enough prompting, right?
- 12:03
And so I think in general, in a broader sense, MCPs are changing how creative tools work. And, um, just an example of that, I was trying to experiment with this a little bit.
- 12:14
We were talking about The client, right, being the orchestrator where it talks to external sources like the API, it talks to local tools like Blender. So what if we extend that anal-analogy a little bit and the client orchestrates between all these different things?
- 12:35
So let's say the user wants to make a game, and people don't come in with the intent of learning Unity, right? They want to make a game, and people don't come in with the intent of learning Ableton, which is a music creation software.
- 12:47
They come in with the intent of making music. So how do we capture that intent and then transform it so that this person doesn't have to worry about the underlying tools, but also it can match their vision?
- 13:01
So I feel like MCPs are going to be a fundamental glue that holds this thing together, and these LLMs are going to be at the center, at the intelligence, uh, at the center of it, right?
- 13:13
So for example, let's say I want to make a game. So I tell it, "Oh, I wanna make a game," and, uh, this, this LLM is all they interface with, right?
- 13:24
And this LLM can then call Blender and make an asset for the game, the right assets. It can then call Unity and make a whole game engine. Unity is a game engine, and so it can put it in, add collisions, add logic, and make a whole game out of it.
- 13:37
It can call the right APIs to get the right assets and animate it and stuff like that. And it can also call Ableton, which will create the soundtrack for the characters, for the game, right?
- 13:49
So you-- this person doesn't really care about everything outside of it. But now that MCP exists, the client can literally just orchestrate between all of these and, uh, choose the right tools and make a game.
- 14:04
So I don't think we're very far away from this, to be honest. And, uh, I made like a really short demo to demonstrate how it could work. So in this one, I'm creating a dragon, and I'm just asking it to give a soundtrack.
- 14:19
So I also created the Ableton MCP, which is what you see on the bottom left, which essentially you can prompt to create music. So I thought, let's try and combine them and see what happens. [electronic music]
- 14:51
So you can see that it creates-- it gets the dragon that I described. It creates some sinister lighting because it's supposed to be a villain, and it also calls Ableton to create a soundtrack for it.
- 15:03
Sure, the quality might not be the best in the world right now, but at the same time, it's-- you can see it's only a matter of time before it starts stitching all these pieces together to create something insane.
- 15:15
So a few questions, honestly, that this brings up for me personally is that, is it just going to be tools talking to each other now, right? Are, are people just going to interface with the LLM and don't, don't have to even worry about learning the complex UI beneath it?
- 15:32
Because that's what seems to be happening, right? It's a whole new group of people that can come in and create now, and they can do it without learning the UI behind it.
- 15:42
And are creatives going to be more like orchestra conductors? Because, because now I think rather than knowing how to play the instrument, it's more important how to get the vision in your head and place it to the LLM so the LLM can execute it for you, and how you're conducting these different pieces that go together.
- 16:01
So it's an interesting time, I think, to be a creator, and MCPs are pretty much right at the center of it. And I think some of these examples like Blender MCP and Ableton MCP are examples of that.
- 16:14
And from that, there have been spurred a lot of MCPs for other creative tools, like I've seen for PostGIS, Houdini, Unity, Unreal Engine. There have been so many. So it's only a matter of time, I feel that, you know, this-- everyone can become a creator, uh, which is pretty cool.
- 16:31
Uh, thank you. Yeah, that's me. And, uh, yeah, you can find me on GitHub [REDACTED:username], and you can find me at Twitter [REDACTED:username]. Um, yeah. So thank you. Thank you for coming to this.
- 16:44
And if you have any thoughts, please leave them in the comments. Thank you.