AI agent for Blender
Retopology, UV rebuilds, mesh map bakes, scene cleanup, naming and export prep. The work that is mechanical, repetitive, and the reason a forty-asset batch takes a fortnight. Mixar puts an agent inside a Blender-based editor so it can read the scene and run that work as one briefed pass.
Mixar is a 3D editor built on Blender, shipped as its own application on Windows, macOS and Linux. The toolset, shortcuts and .blend files are unchanged; the agent is the addition.
The premise
Almost nobody who models for a living is short of ideas. What they are short of is the hours between a finished sculpt and a shippable asset. That gap is a fixed sequence: apply the transform, weld the doubles, strip the loose geometry, recalculate normals, retopologise to a budget, rebuild the UVs the remesh destroyed, bake the high-poly detail down, name everything to the project convention, set the pivots, and export. None of it is creative. All of it has to be correct.
Most AI tooling aimed at 3D targets the other end: generating the asset. That is the part artists are least short of and most protective of. The automation-shaped problem is downstream, it is high volume, and it is where deadlines are actually lost.
What it automates
Each guide below is written to be useful without Mixar installed. They cover the manual technique first, cite the relevant part of the Blender manual where it is the better reference, then say where automation helps and where it does not.
Retopologise a high-poly mesh without leaving Blender: select it, brief the agent, get quads back with UVs rebuilt. And when to still do it by hand.
Read the guide →AI UV unwrapping in Blender versus Smart UV Project and manual seams. Which method to use, how to check texel density, and where automation stops.
Read the guide →Baking normal, AO and curvature maps in Blender: the settings that actually matter, why bakes come back wrong, and how to run the whole pass as one brief.
Read the guide →How an AI texture generator makes base colour, roughness, metallic and normal maps for Blender, and what separates a usable set from a flat one.
Read the guide →Auto rigging in Blender: what a generated skeleton and skin weights get you, the body plans it handles, and the rig work that is still yours afterwards.
Read the guide →AI terrain generation in Blender: what gets built, the controls behind it, how a photo becomes a heightmap, and where it stops short of a terrain package.
Read the guide →A complete Blender scene cleanup checklist: n-gons, doubles, loose geometry, normals, transforms and orphan data, and how to run it as one brief.
Read the guide →Blender batch export to FBX and glTF: the scale, axis, naming and material rules that stop assets breaking in Unreal, Unity and Godot.
Read the guide →How Blender MCP works, how to install it, what it handles well, and the two hard limits an agent running inside the editor removes.
Read the guide →How it works
The agent, called Mixie, runs inside the editor rather than talking to it through a bridge. It reads the live scene: object hierarchy, mesh statistics, transform state, normal orientation, UV layers, material assignments and modifier stacks. It can then act, look at what changed, and correct itself, which is the loop that makes a multi-step pass survive contact with a real file.
The tool surface is deliberately not one flat list. Work is routed to specialists, each seeing only the tools its job needs: modelling, modelling edits, environments, terrain, texturing, UV unwrapping, UV layout, rigging, animation, camera, lighting, rendering and scene administration. A UV task never has to reason past a rigging tool it will not use, which is the practical reason a narrow surface beats a large one.
Toggle it on and the agent drafts the sequence and stops. You approve it, or you tell it what is wrong and it revises. Nothing touches the file until you say so. On a cleanup pass across dozens of objects, that is the whole difference between automation and gambling.
The honest version
Automatic retopology aligns to curvature, and curvature does not know where an elbow bends. On hero characters going into a rig, a human still needs to look at the wireframe and place the loops. Anyone claiming otherwise has not shipped one.
Flattening and packing are solved problems. Deciding where a visible discontinuity is acceptable on this particular asset is not. Expect to mark seams on anything with a defined hero angle and hand off the rest.
A generator produces a plausible, physically correct material. It does not know your project's conventions, which crate is set dressing and which is a hero prop, or what the wear pattern is supposed to say about the object.
Everything else it reaches
Mixar also orchestrates generation from inside the editor, so results land in the scene instead of a download folder: text to image with reference images, depth-guided rendering from a viewport blockout, image to 3D, segmentation of an image into separate meshes, whole-scene assembly from one reference, PBR map generation onto a selected mesh, procedural material generation and auto-rigging. Which engines back each of those is configured server-side and changes, which is exactly why it is worth having them behind one interface instead of seven browser tabs.
It is deliberately the second story. Generated output is a starting point that still needs the finishing pass, and the finishing pass is the thing this cluster is about.
An AI agent for Blender is a system that reads the scene and performs operations in it, rather than answering questions about it. The distinction that matters is scene access: a chat window that writes you a script has to be pasted into Blender and run by hand, while an agent inspects object counts, mesh statistics, transform state, UV presence and material assignments directly, then acts and checks the result. Mixar builds the agent into the editor, so there is no bridge process between the two.
Partly, and the split is consistent. The geometric steps have been automatable for years: QuadriFlow lays out quads, Blender's unwrap solvers flatten islands, Cycles bakes maps. What an agent adds is the judgement and sequencing around those steps, which is where most of the time actually goes: inspecting the input, cleaning what would make the operation fail, picking a target density for the asset, rebuilding UVs afterwards and baking the detail down. Hero assets with strict deformation requirements still want an artist over the result.
Blender MCP bridges an external LLM client to Blender over the Model Context Protocol. It is a genuine breakthrough and it works. The architectural difference is that every operation is a round trip through the protocol, the model only sees what the protocol chooses to expose, and long chains get slow and can hit tool-call timeouts. Mixar runs the agent inside the application against live scene state, and splits its tool surface across specialist agents so each one sees a narrow, relevant set rather than one flat list.
Only if you let it. Plan Mode is a toggle that makes the agent draft a step-by-step plan and wait: you approve it, or you give feedback and it revises before anything executes. On a destructive pass across dozens of objects that approval step is the difference between automation and gambling, which is why it exists as a first-class control rather than a prompt convention.
Open a scene you already have, pick the pass you were about to do by hand, and brief it. Mixar runs on Windows, macOS and Linux, and your Blender shortcuts come with you.