AnyMachine: How an AI Agent Builds and Rigs Machines in Blender

An AI agent that can build a walking machine in Blender — 160 parts, hydraulics on every joint, a 104-bone rig and three animations — is a striking demo. AnyMachine is the open-source workflow behind it, and the interesting part is not the result. It is the method.

Why Agents Usually Fail at 3D

Ask a coding agent to model something in Blender and it writes plausible Python that produces plausible-looking geometry — until you look at it. Proportions drift, parts float, pivots sit in the wrong place. The agent cannot see what it built, so it has no way to notice.

AnyMachine’s answer is structure: force the agent to generate references first, then render its work beside them and correct what does not match.

The Pipeline

  • Reference sheets — one concept image expanded into a turnaround, single views, and a close-up for every joint and mechanism
  • Parts breakdown written from those references
  • One Python script that builds geometry, materials, rig and animations
  • Check renders beside each reference, followed by a second pass on whatever fell short

The script-as-output design is what makes this practical. A Blender file is opaque; a build script can be read, edited and rerun. Change one dimension and regenerate.

Rebuilding Generated Meshes on Real Axes

AnyMachine can also start from an AI-generated mesh, and how it handles that is worth noting. Generated meshes come out as one fused shell — useless for a machine, which needs separate parts that rotate around real pivots.

So the workflow rebuilds the moving parts on proper axes rather than trying to rig the shell. That is the correct call, and it echoes the main problem with every AI 3D generator: generation is cheap, but usable structure is not.

What It Costs

The workflow is MIT licensed and free. Running it is not: you need a capable coding agent, and the reference sheets and close-ups were generated with paid image and video models through fal. Budget for API usage, especially on complex machines where it generates more close-ups.

Mechanical Rigging Without an Agent

If you would rather rig machines by hand, the principles AnyMachine automates are the same ones you would follow manually: separate parts per moving piece, one bone per part, pivots on the true rotation axis, constraints for hydraulics. Our hard-surface guide covers the modelling side.

The Bigger Picture

Agents in Blender are moving from “write me a script” to structured, self-checking workflows. AnyMachine is one of the clearer examples, and the reference-then-verify loop is the pattern worth copying — whatever you are building.

Tool page: AnyMachine. Background on agents in 3D: what Blender MCP is.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top