By now, we’re seeing GPT 6 Astra examples pop up everywhere. So here’s one more. This time, it’s more of a study: can Blender MCP + Astra actually build some cool stuff? Let’s dive in. 1/8

2/8 The starting point: VANTA / 07 - an original (generated image) helmet study. The goal wasn’t just to make a convincing image. It was to see whether Astra + Blender MCP could turn a visual brief into a coherent, editable 3D object under continued human art direction. As always - I like to start with a reference sheet (GPT makes excellent ref sheets) and I even use these on projects I'm building assets for manually - it's an excellent way to start off a 3D build specifically.

3/8 Immediately I had my concerns. It seemed like GPT was cutting corners to get to the finish line. I had to firmly interject several times to have Astra practically start from scratch after the first 2 attempts.

4/8 After a while things started to click. Something to remember - GPT may have MCP access and awareness of all of the tools and what 3D models are, but it's not a 3D modeler or artist. You have to instill process verbally for it to take the time to consider how to approach things from a different prospecting - in this case, use Ortho views to ensure better consistency with the reference.

5/8 One of the biggest issues initially was arbitration. Random primitives, random floating pieces, poor connections, and this wan't fully resolved in the end, BUT it was refined to a much better state that I considered reasonable (at least for the experiment).

6/8 This phase could've been driven way further, but one of the advantages of GPT throughout this experiment, is the generative ability to create textures on-demand. This helps mitigate the need for libraries and additional assets, so some seamless scratches and grunge were low hanging fruit.

7/8 From the ground up I wanted to remain as hands off as possible, so everything including the turntables and staging I simply requested and lived with the results.

8/8 Ultimately, the experiment was interesting. I wasn't sure what to expect, previous MCP asset developments had left me extremely dissatisfied and wasted hours if not days. This was passively over the span of 3 hours and the results are fairly useable to an extent. Let me know your thoughts, and if you have an asset you'd like to see tested - drop it in the comments and I'll see if we can make it happen!

Thanks for doing a detailed breakdown and articulating what it's not good at as well. What I'm most curious about is what is the cost and 'work experience' of dealing with an LLM in place of a human/freelancer? Worth it? Not yet? Humans fucked?

Ya those are all good questions - honestly no clue.. I mean it’s A LOT of tokens and surely this model will eventually be obsolete to the next and so on. But, within the process there were times I just wanted to jump in and do the damn thing myself and show it but I tried my best to take a back seat and just direct as reasonably as I could. Still, in just several hours practically hands off it was impressive to see these results (prior LLMs even the best Claude etc were no where near). But one thing is clear - without direction it was incredibly far off R1 - but then again so are humans at times. I guess it comes down to how well you like hand holding for a couple hours haha.

Interesting. So the model being used (more so than the MCP) is the determining factor for success. And they’re just now becoming capable, so $$ for a while. So no local, open source models will be doing this for a while I suppose.

That’s probably the biggest question - can the local models run with the same awareness and understanding. Sure the same MCP can be accessed anywhere, but it will take some time to reach the level of Astra for throughput. (At this rate time being a matter of months maybe for local models), but could we potentially see a data center running thousands of asset creations daily? Maybe so.

Very interesting case study the topo seem decently clean. But dont do like open Ai spenser tell us token usage and time it took😅

I'll get the approx numbers today, I'm running a few more elevated processes. Surprisingly, it takes the learnings from one exercise and if you use it to build a skill - it really does well carrying that over to the next. You could potentially work on optimizing topo for a day and then lock that that work in as a skill which would be interesting. I'm curious to see how deep the rabbit hole goes

Yeah totally, in my experience with houdini mcp from the last 6 months its getting better and better as time goes.

That's not that bad for the amount of work it did and the result, thanks for posting it 🙏

This is a great study. I'd be tempted to try a similar experiment. Or if you want to try recreating this model I can send it to you

That’s an interesting one - I would force it to build and not provide any geo refs so maybe some views and then some ref plates of the relief and then see what it could do. First pass block out shape and then push details from there. I think it would really struggle with the details though.

Blender MCP plus Astra as a study of whether it can build cool stuff is the 1/8 series I am watching through

This helmet is so cool!

Thanks - I had very little to do with it haha

No, this is related to your personal taste.

These are great. But let me know if Astra can build using Threejs. I’ve built a open source tool for it -

Thanks for sharing! Looks worth exploring for sure

The useful part of this thread is treating it as a study, not another one-shot flex. Did you keep the failed passes?

Ya I’m releasing this and the latest study with all files shortly on GH .. the latest study is pretty interesting and I was surprised so stay tuned

what specific "cool stuff" do you hope to create with MCP and Astra?

Haha who knows - honestly just taking baby steps with a few assets and driving skill sessions from it. A symmetrical space helmet is really nothing too wild I can model this 3X better in the same time, but both learning how to direct it and allowing it to build an understanding at the same time can become valuable in the long run. Not to mention the concurrency - I’m working on multiple apps and plugins on a diff machine with the same API, so elevating productivity with less hands on response is always pretty cool!

the correction loop is the useful part. a helmet that needs three nudges before silhouette and connections settle is closer to real asset work than one-shot prompt theater. Blender MCP still has to survive the cleanup pass.

뭐해먹고 살라고...

If you're an artist - you have the benefit of both sides - servers are down you can cook still, and in parallel you can coach MCP level creation the right way that others can't and 2X your game

Can't you just use Blender with Computer use ?

Absolutely, you can sit down for a couple hours and model exactly what you want, that's perfect for users who model and the right investment imo. But, this does show that someone can get from A-Z (maybe A to T haha) easily enough passively in a couple hours if they know how to communicate what they need. The biggest find here is the hours of learning can wrap up into a skill session that now tossing ALL of the work out - it could replicate almost exactly back to the final check point in a fraction of the time. But it's all still very reliant on human conditioning to get good results.

How much did that experiment cost?

good question - I'll get usage numbers later today.. I did have to burn 2 usage resets BUT I was working on 3 other projects on a separate machine with the same API

That’s great man, thanks!

Do you have the 3D model on GitHub anywhere? I’d like to use it as a use case and create a landing page for it on Vercel. Let me know. Cheers!

Sure, I'll add it later today!

Cool!