Structured Local Automation for Blender

The bridge listens on a local endpoint inside Blender and can be started or stopped from the add-on. It is not a cloud connector, and it does not move the Blender runtime into a hosted service. That boundary keeps the workflow close to the files, assets, add-ons, and render environment the artist or developer already controls.

A terminal command sends a JSON document to the bridge. Blender performs the requested operation and returns a structured result. Codex can compose those commands, but the protocol does not depend on Codex. This makes a health check, a smoke test, or a saved example understandable as an ordinary local development artifact rather than an invisible conversation with the scene.

The agent can help plan an operation, but Blender remains the place where the scene actually exists.

Codex Blender provides named operations for creating starter rooms, outdoor scenes, furniture, lights, procedural primitives, and architecture presets. It can list local assets, search an asset manifest, import supported model formats, inspect scene objects, transform or duplicate named objects, add simple keyframe animation, apply materials, select a render preset, render a PNG, save a blend file, and export GLB or OBJ.

A structured command records both the action and its parameters. That is more useful than a long, opaque block of generated Blender scripting when the operation belongs to a supported workflow. Examples in the repository are validated, reusable commands, so a scene step can be reviewed before it runs and repeated after the project changes.

{
  "action": "create_room",
  "params": {
    "style": "modern_neon"
  }
}

There is a limit we state plainly: the bridge cannot turn any arbitrary image into a perfect production 3D model by itself. Reference-driven modeling is iterative. A useful sequence is to place a reference, create or import a base model, adjust geometry, materials, and camera, render the result, compare it with the reference, and repeat.

The project supports that loop instead of hiding it. Reference images can be placed as textured planes, camera and comparison views can be prepared, and helper scripts can create contact sheets, calculate rough image-difference metrics, and write iteration reports. These tools do not replace artistic judgment. They make each pass easier to inspect and discuss.

The local asset manifest serves a similar purpose. Models, textures, and references can carry paths, tags, source or license notes, previews, and scale hints. Automation becomes more dependable when it can ask what assets exist and how they are meant to be used before changing the scene.

The project layout separates the Blender add-on, the bridge client, MCP tooling, skills, examples, assets, renders, exports, and scenes. It also includes validation and smoke-test scripts that can run before or alongside live Blender checks. Those seams let us improve a command or add a preset without turning the entire integration into one monolithic script.

Structured automation does not remove the exploratory nature of 3D work. It gives repeated operations a clear shape: inspect before editing, name the target, send bounded parameters, render the result, and keep the output in a predictable place. That is the level at which an agent can be genuinely helpful while the creator keeps control of the scene.

Our goal with Codex Blender is practical leverage. Procedural building blocks, asset operations, and repeatable render workflows can become commands, while nuanced modeling and visual decisions remain an iterative collaboration with Blender itself.