Clara is an AI-powered content creation and orchestration system. Clara coordinates multiple GPUs to handle knowledge ingestion, video generation, audio synthesis, and content creation workflows.
Updated December 2025 - Now featuring the latest SOTA models including Qwen2.5-72B, LTX-2, FLUX.1, and Whisper large-v3-turbo.
┌─────────────────────────────────────────────────────────────────┐
│ Clara Brain │
│ (GPU 2 - RTX 3090) │
│ ┌─────────────────────────────────────────────────────────┐ │
│ │ Chainlit UI + Knowledge Base + MCP Tools │ │
│ │ ChromaDB Vector Memory + YouTube Ingestion │ │
│ └─────────────────────────────────────────────────────────┘ │
└──────────────────────────────┬──────────────────────────────────┘
│ Orchestrates
┌──────────────────────┼──────────────────────┐
│ │ │
▼ ▼ ▼
┌───────────────┐ ┌───────────────┐ ┌───────────────────┐
│ Media Node │ │ Audio Node │ │ Compute Cluster │
│ (GPU 0-5090) │ │ (GPU 1-4080S) │ │ (GPU 3,4,5-3090s) │
├───────────────┤ ├───────────────┤ ├───────────────────┤
│ ComfyUI │ │ F5-TTS/XTTS │ │ vLLM │
│ SDXL/FLUX │ │ Whisper v3 │ │ Qwen2.5-72B-AWQ │
│ LTX-2/CogVideo│ │ Voice Clone │ │ Video Processing │
└───────────────┘ └───────────────┘ └───────────────────┘
| GPU | Model | VRAM | Role |
|---|---|---|---|
| 0 | RTX 5090 | 32GB | Video/Image Generation |
| 1 | RTX 4080 Super | 16GB | TTS/STT Audio |
| 2 | RTX 3090 | 24GB | Clara Brain |
| 3,4,5 | 3x RTX 3090 | 72GB | Compute Cluster |
| Task | Model | Details |
|---|---|---|
| LLM | Qwen2.5-72B-Instruct-AWQ | 3-way tensor parallel on compute cluster |
| Video | LTX-2, CogVideoX-5B | Text-to-video generation |
| Image | FLUX.1, SDXL | High-quality image generation |
| TTS | F5-TTS, XTTS-v2 | Multilingual + voice cloning |
| STT | Whisper large-v3-turbo | Fast, accurate transcription |
- Windows 10/11 with WSL2 or Linux
- Docker Desktop with WSL2 backend
- NVIDIA Container Toolkit
- 6 GPUs as described above (or adjust configuration)
# Clone the repository
git clone https://github.com/yourusername/clara.git
cd clara/Clara
# Run setup script (Linux/macOS)
chmod +x scripts/setup.sh
./scripts/setup.sh
# Or on Windows
scripts\setup.bat
# Edit configuration
# Add your Hugging Face token to .env
nano .env
# Start Clara
docker compose up -d| Service | URL | Description |
|---|---|---|
| Clara UI | http://localhost:8000 | Main Chainlit interface |
| ComfyUI | http://localhost:8188 | Direct ComfyUI access |
| Media API | http://localhost:8020 | Media node API |
| Audio API | http://localhost:8030 | Audio node API |
| Compute API | http://localhost:8010 | Compute cluster API |
| ChromaDB | http://localhost:8500 | Vector database |
- YouTube transcript extraction and compression
- Document processing (PDF, DOCX, MD)
- ChromaDB vector storage for semantic retrieval
- Video: LTX-2, CogVideoX via ComfyUI
- Images: SDXL, FLUX.1, custom workflows
- Audio: F5-TTS, XTTS-v2, voice cloning
- Podcasts: NotebookLM-style discussions
Clara has access to various tools through the MCP protocol:
- GPU task dispatching
- Knowledge base search
- File management
- Content generation
/ingest <youtube_url> - Add video to knowledge base
/status - Check system status
/memory - Browse knowledge base
/generate - Start content generation
Create a Summary Video:
1. /ingest https://youtube.com/watch?v=...
2. "Create a 5-minute summary of this video with visual highlights"
3. Clara coordinates:
- Compute cluster generates script
- Audio node creates voiceover
- Media node generates visuals
- Compute cluster edits final video
Build Expertise:
1. /ingest [multiple videos on a topic]
2. "What are the key insights about [topic]?"
3. Clara retrieves from knowledge base and synthesizes
Generate Content:
"Generate a cinematic video of a sunset over mountains"
"Create a voiceover reading this script in a calm tone"
"Make a thumbnail image for my video about AI"
Key settings in .env:
HF_TOKEN=your_token_here # Hugging Face access (required)
VLLM_MODEL=Qwen/Qwen2.5-72B-Instruct-AWQ # LLM model
VLLM_TENSOR_PARALLEL_SIZE=3 # GPUs for compute cluster
WHISPER_MODEL=large-v3-turbo # STT model
TTS_MODEL=xtts_v2 # TTS modelIf your GPU indices differ, update docker-compose.yml:
deploy:
resources:
reservations:
devices:
- device_ids: ['YOUR_GPU_INDEX']Clara/
├── docker-compose.yml
├── containers/
│ ├── clara-brain/ # Main orchestrator
│ │ ├── app/
│ │ │ ├── clara_core.py
│ │ │ ├── memory/
│ │ │ ├── mcp_tools/
│ │ │ └── services/
│ │ ├── chainlit_app.py
│ │ └── Dockerfile
│ ├── clara-media/ # ComfyUI node
│ │ ├── workflows/ # JSON workflow files
│ │ └── Dockerfile
│ ├── clara-audio/ # TTS/STT node
│ │ └── Dockerfile
│ └── clara-compute/ # vLLM cluster
│ └── Dockerfile
├── shared/
│ ├── knowledge/ # Ingested content
│ ├── outputs/ # Generated content
│ └── models/ # Model cache
└── scripts/
├── setup.sh
├── setup.bat
├── check_gpus.py
└── ingest_youtube.py
Create JSON workflow files in containers/clara-media/workflows/:
{
"1": {
"class_type": "CheckpointLoaderSimple",
"inputs": {"ckpt_name": "your_model.safetensors"}
},
...
}# Check NVIDIA driver
nvidia-smi
# Verify Docker GPU support
docker run --rm --gpus all nvidia/cuda:12.1-base nvidia-smi
# Run GPU check script
python scripts/check_gpus.py- Reduce batch sizes in workflows
- Use quantized models (AWQ/GPTQ)
- Clear model cache:
docker compose down -v
# Check logs
docker compose logs clara-brain
# Rebuild
docker compose build --no-cache clara-brain# Check network
docker network ls
docker network inspect clara_clara-network
# Restart all services
docker compose restartMIT License - see LICENSE file.