I have been dragging my feet on a side project for two years. I want to make a series of short explainer videos about how Unix tools work, and the missing piece has always been the same thing: I do not want to draw the diagrams by hand, and I do not want to pay a SaaS subscription to generate them. Blender is the obvious answer (it is free, it is open source, and it can render almost anything), but every time I have sat down to learn it, I have ended up in a rabbit hole of viewport shortcuts and node graph tutorials. Blender is the most powerful 3D tool most people have never finished learning.
This month I tried something different. I asked a coding agent to drive Blender for me, and it worked on the first try. The result is not a complete explainer video, but it is a usable render of “a pelican riding a bicycle,” produced from three prompts and zero Blender expertise. If you have been waiting for the same reason I have, this is the shortest path I have found.
Why this is different from the LLM image tools
The big image generation tools (DALL-E, Midjourney, Stable Diffusion) all produce images directly. Blender is a 3D modeling and animation suite (software that lets you build and animate three-dimensional scenes), and the way to get an image out of it is to construct a 3D scene, set up lights and a camera, and render the scene to a 2D image. The 3D model file (a .blend file) is editable in Blender, and the rendered image can be tweaked, re-rendered, or animated into a movie by chaining many renders together with ffmpeg (a free command-line video tool).
This is the part that has always made Blender feel like a lot of work. The minimum viable use case (a single rendered image of something) requires understanding meshes, materials, lights, cameras, and the render engine. The default Blender UI assumes you know what all of those mean.
What changed with the coding agents is that they already know the Blender Python API (the scripting interface Blender exposes for automation). When you ask an LLM (a large language model, the type of AI behind ChatGPT and similar tools) to “render a scene of a pelican riding a bicycle,” it does not try to draw a pelican directly. It writes a Python script that constructs the scene using Blender’s primitives (basic geometric shapes like spheres, cones, and curves), applies materials and lights, sets up a camera, and calls the renderer. The output is a real 3D model you can open in Blender and edit, not a fixed image you can only regenerate.
What it looks like in practice
I tried this in ChatGPT Codex on a Mac running the latest Blender 4.5 LTS. The setup took about five minutes: install Blender from blender.org (not from the Mac App Store, because the App Store version runs in a sandbox that breaks the Python API), launch it once to accept the license, and confirm that you can run Blender from a terminal by typing /Applications/Blender.app/Contents/MacOS/Blender (the underlying executable that the app bundles).
From there, the workflow is just prompts. The first one I tried was “Use the already installed /Applications/Blender to render a scene of a pelican riding a bicycle.” The agent wrote a Python script that built a stylized pelican, gave it a body that fit on a bicycle, positioned everything, set up a soft daylight rig, and rendered at 1080p. The first render came back in about 90 seconds.
The follow-up prompts were where it got interesting. “OK add a background and a lot of flair” caused the agent to add a sky gradient, some atmospheric haze, and a second pass of detail lighting. “OK make it a whole lot better” caused the agent to redo the materials, add some surface variation, and re-render. Each iteration was a few seconds of compute and a few seconds of waiting. Total wall-clock time to go from “render a pelican on a bicycle” to “I would actually use this as a video thumbnail” was about ten minutes.
The output is a real Blender file with a real mesh, real materials, and a real camera. If I open the file in Blender, I can see the pelican as separate parts (body, beak, wings) that I can move, scale, or re-color. I can change the camera angle. I can add an animation. None of that is locked in.
Why this works
The reason the coding agents are good at Blender is the same reason they are good at any software with a Python API: they have seen a lot of example code in their training data, and Blender’s Python API has been documented online since 2008. The agent does not need to understand 3D rendering to write a Blender script. It needs to know what Blender primitives to use, what API calls to make, and what order to make them in. That information is in the training data.
This is not a permanent advantage. The Blender Python API changes between versions, and Blender 5.0 (currently in beta) has reorganized several core APIs. As of late 2026, the major coding agents still produce code that works against Blender 4.x out of the box, and Blender 5.x compatibility is improving but not perfect. If you are starting a project, pin to a specific Blender version and confirm that the agent’s first render worked before you invest hours.
A second limit is that coding agents are not 3D artists. They will produce renders that look technically correct but artistically generic. The pelican on the bicycle looks like a stock illustration. If you want a specific style (Pixar, anime, retro), you have to prompt for it explicitly, and even then the agent will need a few rounds of iteration to get there. For diagrams, technical illustrations, and explainer-video assets, the default output is fine. For hero shots, you will need either a skilled prompt writer or a human 3D artist.
How to try it yourself
If you have a Mac and 30 minutes, this is the shortest path I have found:
- Install Blender from blender.org. Not the App Store version. The App Store version runs in a sandbox that blocks the Python API. The blender.org installer does not.
- Launch Blender once. Accept the license, close the splash screen, and quit. This creates the user preferences directory the agent will need.
- Confirm the agent can run Blender. In a fresh terminal, run
/Applications/Blender.app/Contents/MacOS/Blender --version. If you get a version string back, the agent can drive Blender. If you get a permission error, the agent needs full disk access in System Settings. - Pick a small first project. Render a single object, not a full scene. The first round teaches the agent your local setup; the second round is where you get useful output.
- Iterate on the prompt, not the file. If the render is wrong, ask the agent to redo it. The script is the artifact, not the file.
Trade-offs
The coding-agent-driven Blender workflow is not free. The first render of a new scene typically takes 90 seconds to 3 minutes, depending on the complexity and the resolution. For an explainer video with 30 scenes, that is roughly an hour of compute, and you will want at least two rounds of iteration per scene, so the realistic budget for a polished 30-scene video is two to three hours of wait time.
A second trade-off is reproducibility. The coding agent does not version the Blender scripts it writes by default. If you want to be able to regenerate a scene six months later, you have to ask the agent to save the script somewhere, or copy it out of the conversation. Most coding agents will do this on request, but you have to remember to ask.
A third trade-off is that the agent does not always succeed on the first try. Blender’s API is large, and the agent occasionally hits an edge case where its generated script produces an empty scene or a render error. The fix is usually a short follow-up prompt that names the error, and the agent will regenerate the script with the fix. In my experience, the failure rate is around one in ten first renders, and the failures are recoverable in a single round.
For anyone who has been putting off a 3D project because the learning curve on Blender feels too steep, this is the easiest entry point I have seen. For anyone who already knows Blender well, the workflow is faster but you will spend most of your time correcting the agent’s instincts rather than steering them.
Bottom line: Coding agents can drive Blender from a chat prompt, the result is a real Blender file you can edit, and the setup is about 30 minutes on a Mac. If you have been waiting for a reason to revisit Blender, this is the reason.