Flint 3D Model Generation Server

FastAPI + MCP stack that turns text prompts into Blender scenes and exports USDZ files via export_model.py.

Overview

• Runs a FastAPI server (api_server.py) with /run, /run-high, /download, and /health endpoints.
• Uses MCP to drive Blender through the blender-mcp toolchain and exports via the bundled Python script.
• Keeps context usage in check with tool allowlists, step caps, and prompt truncation.
• Produces .blend and .usdz artifacts in exports/ for downstream consumption.

Highlights

• 🤖 Claude Sonnet 4.5 integration tuned for MCP workflows
• 🎨 Automated Blender pipeline that saves .blend files and exports USDZ
• 🔄 Context-safe MCP bridge (tool allowlists, token caps, step limits)

Prerequisites

• macOS with Blender available at /Applications/Blender.app/Contents/MacOS/Blender
• Python 3.10+
• uv toolchain (needed for uvx blender-mcp): brew install uv
• Valid Anthropic API key stored in .env (ANTHROPIC_API_KEY=...)
• Recommended: create and activate the project virtual environment (python -m venv .venv then source .venv/bin/activate)

Step-by-Step Setup

1. Install Python 3.10+ and Blender – Blender must be callable at /Applications/Blender.app/Contents/MacOS/Blender or you need to update BLENDER_PATH in api_server.py.
2. Install uv (MCP helper) – brew install uv so the uvx blender-mcp command in api_server.py works.
3. Create a virtual environment – python -m venv .venv && source .venv/bin/activate (recommended but optional).
4. Install dependencies – pip install -r requirements.txt.
5. Create .env in the project root with the secrets required by the scripts:

   cp config.env .env  # optional helper if you track defaults there
   

   Set at minimum:

   ANTHROPIC_API_KEY=sk-your-key
   BLENDER_PATH=/Applications/Blender.app/Contents/MacOS/Blender  # override if needed
   EXPORT_FOLDER=exports  # optional; defaults to project root
   

6. Start the server – python start_server.py (hot reload enabled).
7. Check health – curl http://localhost:8000/health should report healthy once Blender, the export folder, and the Claude API are accessible.

API Endpoints

| Method | Path | Description | |--------|---------------|-------------| | GET | /health | Reports Blender, Claude, and export-folder status | | GET | /docs | FastAPI interactive docs | | POST | /run | Standard-quality MCP generation (token-safe) | | POST | /run-high | High-quality MCP generation (more steps/detail) | | GET | /download/{filename} | Fetch the latest USDZ export |

Example Requests

Standard profile (balanced, ≤5 tool calls):

curl -X POST http://localhost:8000/run \
     -H "Content-Type: application/json" \
     -d '{"prompt": "Create a sci-fi crate with glowing edges"}'

High-detail profile (≤8 tool calls, richer lighting/material cues):

curl -X POST http://localhost:8000/run-high \
     -H "Content-Type: application/json" \
     -d '{"prompt": "Build a stylized medieval watchtower with torches and banners"}'

Quality Profiles & Context Guards

| Profile | Steps | LLM Tokens | Notes | |-----------|-------|------------|-------| | standard | 5 | 1024 | Uses a minimal tool allowlist and aggressive prompt truncation for reliability | | high | 8 | 2048 | Adds richer geometry/material hints and allows the set_texture tool |

Both modes enforce:

• toolContextSize="small" to shrink tool schema payloads
• Explicit tool allowlists so unused MCP connectors never enter the prompt
• Constraint blocks injected into the LLM prompt (max tool calls, snippet size, mandatory save_file)
• Automatic .blend validation plus USDZ export via export_model.py

Switch to /run-high when you need more detail, then fall back to /run if you hit context limits.

... View full README on GitHub