Vork vs Tesseract: 3D Engines for AI Agents Tested on a Mac

By Beau Johnson··8 min read

Quick answer

Vork (from Voyager) and Mirage's Tesseract are 3D engines built for AI agents like Claude Code. Vork is good for simulation-heavy hero shots but slow on my M4 Mac (462 seconds for a 4 second shot vs 7.8 seconds in three.js). Tesseract suits finished social videos if you fit its license terms.

Two 3D engines built for AI agents launched on the same morning. Vork from Voyager showed a million-bird demo, and twenty minutes later Mirage showed a rainy 1980s kitchen that Claude Code built in its Tesseract engine. In this video I explain how each one works, compare the speed claims with what their own fine print says, and test Vork on my M4 Mac against the three.js pipeline I use every day. If you make content with agents and want to know whether these tools are worth trying, this one is for you.

What I cover in this video

  • 0:00 Claude turns one flat AI picture of a kitchen into a 3D room
  • 0:46 The timeline: two engines for agents launch 20 minutes apart
  • 3:02 Why tools like Blender are hard for an AI to drive
  • 3:48 How Vork actually works
  • 6:05 The million birds demo and what the README says about speed
  • 7:37 How Claude built the kitchen in Tesseract
  • 10:41 My Vork test on an M4 Mac vs three.js
  • 12:12 The verdict: who should use what

Why build a 3D engine for agents at all

Blender is free, open source, and great. But Blender, Houdini, Unreal, and Nuke were all designed for a person sitting at a screen, with panels, sliders, node graphs, and menus inside of menus. Agents can already drive Blender today through a community plugin called MCP for Blender, which lets Claude run Python inside a live Blender session. I have it installed myself. The catch is that the agent has to learn a cockpit built for hands before it can make anything.

Voyager's argument comes down to one idea that I think is smart: tools built for humans need thousands of settings because a person can't write a shader in the middle of a shot. An agent can. So you give the agent a small engine and let it write the rest as code. If you're newer to how agents work, my AI Agents page is a good primer.

How Vork works

The loop is simple:

  1. The agent writes a scene. A scene is a JSON text file. The first example in the README is nine lines: a sphere, a floor, a light, a camera, and a render size.
  2. It renders a preview.
  3. The agent looks at the picture.
  4. It fixes the code and goes again.

When a shot needs something unusual, the agent writes a program for it. Materials, cameras, light, simulation, and compositing can all be programs written in a GPU language called Slang. In their black hole demo, Voyager says the agent wrote the black hole physics as a program and plugged it into the engine's light hook, with no black hole code in the engine itself. They're upfront that it's an approximation, not an exact physics solver.

Under the hood the core is written in Rust and renders on your GPU through Metal on a Mac or Vulkan on NVIDIA cards. Fun detail: the camera, the AgX color look, and the lights all follow Blender's conventions. It's built to replace Blender for agents, and it still speaks Blender.

Voyager says the getting-started help is about 2,500 tokens. I saved the v0.1 help page and measured it, and it does come out to about 2,500. Keep in mind that's only the first page, and there are 25 help topics behind it.

The speed claims vs the fine print

The million birds demo runs the flock as compute programs on the GPU. Space is split into a 32 by 32 by 32 grid, and each bird picks seven neighbors to follow. Close birds are real 3D meshes, far ones are flat silhouettes, and the camera plus nine foreground birds were set by hand.

Voyager says it rendered in 97 seconds on one NVIDIA L4. Per the README, that's the render of the 8 second movie only, after about 7 more minutes of setup and warmup, so roughly 8.45 minutes total, on the draft setting. Their site compares that to a projected 13 to 18 minutes in Blender EEVEE. That number is projected, not measured.

The README is more honest than the launch post, and I respect them for it:

These are timings, not equal-quality speedups.

At final quality their own table puts Vork at about 1.1x to 2.3x faster than EEVEE, with different settings on each side, and their path tracer at about 2x to 5.8x slower than Cycles. For "up to 24x faster" and "2x more token-efficient" from the launch post, I couldn't find a published method for either. It's a real engine doing real work. The post has the best case numbers and the README has the honest ones.

How Claude built the kitchen in Tesseract

This is the one with Claude's name on it. Mirage's recipe:

  1. Reference: one still from GPT Image 2.6 of a messy 80s kitchen on a rainy night.
  2. Camera: Claude Code solved the camera from the vanishing points, landing on a 21 mm lens and an eye height of 134 cm.
  3. Build: Claude wrote scripts that built the room with AI-made textures.
  4. Review: two critic agents, an art director and a specialist, compared each round to the reference. Their public 3D skill keeps fixing until both score it at least 8.5 out of 10, for up to three rounds.

It was path traced at 128 samples per pixel in 1080p, about 45 minutes for 312 frames on one RTX PRO 6000. Mirage also told you the cheats: the yard is a relit 2D backplate, and the rain and creature are 2D cards. Movies have done that for a long time, so it doesn't bother me. I didn't test Tesseract, because its terms count testing as commercial use for businesses making $1 million a year or more, and my business is over that line.

My Vork test on an M4 Mac

I ran Vork 0.1 from the official GitHub release in a sandbox on my M4 Mac (macOS 26.5.1). I had Claude write a context window shot: clay token tiles on a ceramic tray, tiles past the edge grayed out, the newest one glowing, with a slow camera dolly. Claude wrote it as a script of about 40 lines that outputs the scene JSON.

  • First-time shader setup: 111 seconds, once
  • Simple 1080p still: about 5.6 seconds
  • Their 2 second product orbit example: 103 seconds
  • My 4 second shot at final quality: 462 seconds (4.8 seconds a frame)
  • Same shot in my three.js pipeline: 7.8 seconds

They're different renderers with a different look. Vork looks like a product photo, with reflections and soft shadows. three.js looks like a clean motion graphic. You're trading speed for the look.

Key takeaways

  • For simulations like a flock or a black hole, try Vork. It's source-available, free during the research preview including commercial work, and what you make is yours, but you can't resell it or run it as a hosted service.
  • For a finished social video with 3D, motion graphics, and sound in one project, Tesseract is built for that if you fit its terms.
  • If you already live in Blender, MCP for Blender works today. The agent just has to learn a lot of menus.
  • three.js stays my daily driver because I post every day. Vork goes in the toolbox for hero shots.
  • Unlike a video model guessing pixels, these engines build a world you can relight, move through, and render again.

FAQ

Can Claude Code make 3D scenes and animations?

Yes. Mirage showed Claude Code building a 3D kitchen in its Tesseract engine from one reference image, and I had Claude write a Vork scene as a script of about 40 lines that outputs scene JSON.

Is Vork open source?

No. Vork is source-available, not open source. It's free during the research preview, including commercial work, but you can't resell it or run it as a hosted service, and later versions may be paid.

How fast is Vork on a Mac?

On my M4 Mac, a simple 1080p still took about 5.6 seconds, but my 4 second shot at final quality took 462 seconds. The same shot in my three.js pipeline took 7.8 seconds, with a different look.

Is Mirage's Tesseract free to use?

Tesseract is a free local engine, but it isn't open source. Commercial use is free unless your business makes $1 million a year or more, and its terms count testing as commercial use.

Chapters

  1. 0:00Claude built a 3D kitchen from one picture
  2. 0:44Two engines, one morning
  3. 2:24Shipping Skool
  4. 2:40Why Blender fights an AI
  5. 3:53How Vork actually works
  6. 6:01A million birds, and the fine print
  7. 7:58How Claude built the kitchen in Tesseract
  8. 10:33I tested Vork on my Mac
  9. 12:32The verdict: who should use what

Transcript

Read the full transcript

0:00Claude just took one flat AI picture of a rainy 1980s kitchen and turned it into a 3D room you can move a camera through. And according to Mirage, the company behind the engine, not one frame of it came from a video model. You've got lightning flashing, a microwave counting down, and a lamp that goes out right on Q. 20 minutes earlier, a different team dropped an engine built for agents that rendered a million birds, each one steering itself. So, that's two 3D engines made for AI agents instead of humans, both showing off on the same morning. And in this video, we're going to be breaking down how each one actually works, what the fine print says, and I installed one of them on my own Mac to see if it lives up to the hype. Let's jump right into it. So, let's start with the timeline because

0:46this happened fast. On November 6th at 8:38 in the morning Pacific, Ravi Periq posted this. Dear Claude, you are being held back by human- ccentric graphics APIs. And then he says he built a 3D NVFX engine where agents can write their own shaders, materials, cameras, and physics. And the core API fits in about 2,500 tokens. It's called Vorc. He also claims it's two times more token efficient than Blender, Unreal, and Houdini, and up to 24 times faster at rendering. Now, those are his claims, and we're going to check them later in this video because his own readme tells a slightly different story. So, the post blew up. 467,000

1:31impressions and 6,800 bookmarks by the next day, which is a crazy bookmark number. To me, that says people want to actually try this thing. The repo went from 742 GitHub stars to about 1,100 a day later, and version 0.2 2 shipped the day after launch. So then 20 minutes after Ravi's post, Mirage posted the kitchen. Mirage is the company behind Captions, the video editing app, and their post said, "No video models were used in the making of this scene." Their engine is called Tesseract. They announced it back in September as in their CEO's words, Premiere and After Effects rebuilt for agents. And earlier this month, they added a 3D skill to it. And by the sixth, path tracing. So,

2:16you've got two teams betting on the same idea. The next user of 3D software is going to be an AI agent. Real quick, every video on this channel is made in code with Claude Code. Inside Shipping School Premium, I show you how to turn your own product into videos that get views step by step. Plus, a live Q&A with me every Saturday. Join free link below. So, why would anybody need a new 3D engine? Blender is free. It's open- source. And it's amazing. Here's something most people don't think about. Blender, Houdini, Unreal, and Nuke were all designed for a human sitting in front of a screen. You've got panels, sliders, node graphs, and menus inside of menus. Voyager, the team behind Vorc,

3:02puts it like this. Blender, Houdini, Unreal, and Nuke were built for humans. Vor is built for agents. So, agents can already drive Blender today. There's a community plugin called MCP for Blender. It has over 33,000 stars on GitHub and it lets Claude run Python inside of a live Blender. I actually have it installed on my machine. But think about what the agent is doing there. It's operating a cockpit that was built for hands and it has to learn all of those controls before it can make anything. Voyager's whole argument comes down to one line and I think it's a smart one. Tools built for humans need thousands of settings because a person can't write a shader in the middle of a shot. An agent can. So instead of handing the agent a

3:48giant control panel, you give it a small engine and let it write the rest as code. So how does VORC actually work? Don't get lost in the sauce here. It's pretty simple. Step one, the agent writes a scene. A scene is just a text file in JSON. The very first scene in their readme is nine lines long and it's a sphere, a floor, a light, a camera, and a render size. Step two, it renders a preview. Step three, the agent looks at the picture. And step four, it fixes the code, and goes again. The readme says it straight. An agent writes a scene, renders it, looks at the result, and revises the code. There is no editor to drive. Now, here's what this is all about. When a shot needs something unusual, the agent writes a program for

4:34it. Materials, cameras, light, simulation, and compositing can all be programs written in a GPU language called Slang. You can see it in their black hole demo. Voyager says the agent wrote the physics of the black hole as a program and plugged it into the engine's light hook with no black hole code in the engine. So, the light bends because the agent wrote the code that bends it, which is insane. And to be fair, they say it's an approximation, not an exact physics solver. So under the hood, the core is written in Rust and it renders on your GPU through Metal on a Mac or Vulcan on Nvidia cards. By default, it mixes regular rasterizing with hardware ray trace lighting. And there's a path tracer, too, if you want it. So remember that 2500 token number. We checked it.

5:20Their site says the getting started help is about 2,500 tokens. I saved that help page from version 0.1 and measured it and it comes out to right around 2500. So that part holds up, but that's just the first page. And there are 25 help topics behind it. So here's a funny detail. The camera follows Blender's physical camera. The color follows the same AGX look Blender uses, and the lights follow Blender's lights, too. So it's built to replace Blender for agents, and it still speaks Blender. GitHub even lists Claude as a contributor on the repo, but Voyager credits its big showcase shots to Astra and GPT 6.1 Soul, not Claude. So, let's talk about the demo everyone shared. A

6:05million birds over a lagoon, each one steering itself. Here's how it actually works. The flock runs as compute programs on the GPU. space gets split into a 32x 32 by 32 grid and each bird picks seven neighbors to follow from a short list of birds close by. Birds near the camera are real 3D meshes and the far ones are flat silhouettes. The camera and nine birds in the foreground were set by hand. And fun fact, the million don't actually flap. 1 million is also the cap because that's the most their bird IDs can hold. So Voyager says it was simulated and rendered in 97 seconds on one Nvidia L4. And per their readme, that's just the render of the 8-second movie, but their readme shows about 7 more minutes of setup and

6:51warm-up before that for about 8 1/2 minutes total. And that run used their draft setting. So now the Blender number. Ravi's post said Blender's Eevee would take 31 minutes. That was a projection, not an actual Blender render, and it was at four times the samples. their own site now says a projected 13 to 18 minutes about 8 to 11 times longer projected not measured. So then there's the speed section of the readme and honestly respect to them for writing this quote these are timings not equal quality speedups at final quality their own table has vorc at about 1.1 to 2.3 times faster than Eevee with different settings on each side. Their path tracer is two to almost six times slower than Blender cycles. And they say

7:37independent reproduction of the comparisons isn't working right now. As for 24 times faster and two times more token efficient, I couldn't find a method for either one anywhere. Take that for what you will. So, is it all hype? No. It's a real engine doing real work. The post has the best case numbers and the readme has the honest ones. So, now the kitchen, and this is the one with Claude's name on it. Mirage posted a thread on exactly how they built it, and it's a really clean recipe. Step one, the reference. One still image from GPT image 2.6, a messy '8s kitchen on a rainy night with a full moon. Step two, Claude Code solved the camera from the vanishing points in that picture. It worked out a 21 mm lens and an eye

8:23height of 134 cm so the 3D room would line up with the image. Step three, according to their thread, Claude wrote scripts that built the room with AI made textures. Step four, two critic agents compared each round against the reference. And their public 3D skill has that same idea built in. An art director and a specialist. And the agent keeps fixing until both score it at least 8.5 out of 10 for up to three rounds. Then they animated it. A slow pushin. The microwave turntable spins while the display counts down. And at zero, it beeps three times and the lamp goes out. The lightning is real light in the scene. And in the close lightning strike, a creature stands in the yard. The rain, thunder, and microwave hum are

9:0911 Labs sound effects, and the thunder is timed to how far away each flash is, which is such a cool detail. And then the render path traced at 128 samples per pixel in 1080p, about 45 minutes for 312 frames on one RTX Pro 6000. That's around 13 seconds of video, so roughly 9 seconds a frame. So, here's what I love about this thread. They told you the cheats. In their words, the yard is a relit 2D back plate and the rain and the creature are 2D cards. The shot opens as gray clay and then a wipe reveals the render to prove the room really is 3D. And the whole shot is one editable file, so you can move the camera and render it again. Movies have been sneaking flat

9:55paintings in behind real sets for a long time, so I don't mind it at all. So, yes, it's public. It's a free local engine for Mac, Windows, and Linux, and you can add it to Cloud Code with one command. npx skills add mirage-huract. Now, the terms, it's not open- source, and commercial use is free unless your business makes $1 million a year or more, and then you need a deal with Mirage. Their terms also count testing as commercial use. So, I didn't test Tesseract for this video. Our resale business is over that line. So, for now, that is the dealbreaker for me. So, I installed Vorc on my own Mac. It's an M4 on Mac OS 26.5. I used version 0.1 from

10:41the official GitHub release. Check the check sum and ran it in a sandbox, never on my main system. Vorc needs Apple Silicon on Mac OS 26 or a recent Nvidia card with hardware ray tracing. And funny enough, the big data center AI chips like the H100 can't run it because they don't have the ray tracing hardware it needs. I had Claude write one shot and I picked something I explain all the time on this channel, the context window. Clay token tiles sit on a ceramic tray with a dark metal edge where the window ends. The tiles past the edge sit outside, grayed out, and the newest one glows. Then a slow camera dolly with a little depth of field. Claude wrote it as a script of about 40 lines that spits out the scene JSON. So, here are the real numbers on my Mac. The

11:26first time shader setup took 111 seconds, and that only happens once. A simple 1080p still took about 5.6 seconds. Their product orbit example, 2 seconds of video, took 103 seconds. And our 4second shot took 462 seconds at final quality. That's 4.8 seconds a frame. So, a 15-second clip would take about half an hour. Yes, it's slow, but what am I in a hurry for? Well, I make a video every single day, so I guess I am kind of in a hurry. So, then I had Claude build the exact same shot in the 3.js pipeline we use for every 3D clip on this channel. Same tray, same tiles, same camera move. That one rendered in 7.8 seconds on the same Mac. That's

12:12about 60 times faster. So, look at them side by side. You can see Vor's metal edge actually reflects the glowing tile next to it. The glow spills onto the tray and the shadows are soft. Ours looks like a really clean motion graphic, and Vorc looks like a product photo. So, you're trading speed for the look. Which one looks better to you? So, who should use what? If your shot is a simulation, like a flock or a black hole, Vorc is the one to try. And I think the code first idea is the right one for agents. It's source available, not open source. It's free during the research preview, including commercial work. And what you make is yours. You just can't resell it or run it as a hosted service. And later versions may be paid. If you want a finished social

12:57video with some 3D inside of it with the video, motion graphics, sound, and 3D allinone project, Tesseract is built for that as long as you fit their terms. And if you already live in Blender, Blender is free and open source. MCP for Blender is a free community plug-in and it works today. The agent just has to learn a lot of menus first. So for us, 3.js stays our daily driver because we put out a video every day and the speed matters. Vorc goes in the toolbox as a backup for the hero shots where the look matters more than the clock. So I think the bigger deal is what these engines actually make. A video model guesses pixels. These build a world you can relight, move the camera through, and render again. Even with a couple of 2D

13:42cards sneaking in, that's huge for anyone making content with agents. So, what about you? Would you hand your 3D work to an agent, or is Blender staying open on your desktop for good? Let me hear it down below. And if you want more breakdowns like this one, give it a like so YouTube shows it to more builders. Have a blessed day.

Ready to start building with AI?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

Keep reading