Content
# LL3M Agent (Monorepo)
A unified autonomous 3D modeling system for Blender, built on the Model Context Protocol (MCP).
[](https://www.youtube.com/watch?v=BnbDg1-be4g)
This project integrates the multi-agent reasoning framework from the LL3M research with a local Blender execution bridge. It provides a production-ready stack for generating 3D assets, inspecting scene hierarchies, and performing iterative refinements via natural language.
## Architecture and Origins
The system is derived from two foundational concepts:
- **LL3M (Large Language 3D Modelers)**: A multi-agent methodology for 3D generation through interpretable Python code. [Source](https://github.com/threedle/ll3m)
- **Blender Lab MCP**: The official vision for exposing Blender as a tool for large language models. [Documentation](https://www.blender.org/lab/mcp-server/)
## System Capabilities (v3.1)
### Autonomous Modeling Pipeline
The system implements a structured multi-agent loop to ensure geometric and physical accuracy:
- **Planner**: Decomposes natural language requests into discrete geometric components and spatial requirements.
- **Writer**: Translates modeling plans into modular Python scripts utilizing `bpy` and `bmesh`.
- **Debugger**: Analyzes execution tracebacks and visual feedback to perform automated error correction.
### Local Retrieval-Augmented Generation (RAG)
Integrated Python-based RST parser that provides agents with:
- Fully qualified API signatures for `bpy` and `bmesh`.
- Contextual usage examples extracted from official documentation.
- Automatic truncation and navigation for large module references.
### Scene and Object Intelligence
- **Hierarchical Summarization**: Detailed mapping of collections, object visibility, and active selection states.
- **Technical Inspection**: Direct access to mesh data, material node trees, and modifier stacks.
- **Visual Feedback**: Real-time viewport screenshot capture and high-resolution background rendering.
## Repository Structure
- `/brain`: Node.js/TypeScript MCP server handling agent orchestration and documentation retrieval.
- `/body`: Python environment containing the Blender bridge, modeling tools, and full API RAG dataset.
- `/skill`: Gemini CLI skill definition for managing the autonomous workflow.
## Installation and Setup
### 1. MCP Configuration
Register the server in your configuration (e.g., `~/.claude.json`):
```json
"ll3m": {
"command": "node",
"args": ["/path/to/ll3m-agent/brain/dist/index.js"]
}
```
### 2. Blender Add-on
- Install the add-on located in `body/addon/blender_mcp_addon/`.
- Enable "Start Server" within the add-on preferences (Default Port: 9876).
## Operation
Execute the modeling loop by providing a technical or descriptive request:
> "Use the ll3m-agent skill to generate a minimalist industrial interior with a glass-top table."
The system will proceed through the PLAN → RETRIEVE → WRITE → EXECUTE cycle automatically.
