Getting Started with NemoClaw

NemoClaw is NVIDIA's enterprise AI coding tool optimized for GPU-accelerated inference.

Prerequisites

- NVIDIA GPU with 16GB+ VRAM (RTX 4070 Ti or higher)

- NVIDIA Driver 535+ installed

- Docker with NVIDIA Container Toolkit

Installation

```bash

Pull the NemoClaw container

docker pull nvcr.io/nvidia/nemoclaw:latest

Run with GPU access

docker run --gpus all -it \

-v $(pwd):/workspace \

-p 8080:8080 \

nvcr.io/nvidia/nemoclaw:latest

```

Or install via pip (lighter weight)

```bash

pip install nemoclaw

nemoclaw init # Downloads default model

```

Configuration

```yaml

nemoclaw.yaml

model:

backend: tensorrt-llm

name: nemotron-coder-22b

quantization: int8

inference:

max_tokens: 4096

temperature: 0.1

gpu_memory_fraction: 0.9

security:

audit_log: true

allowed_paths:

- /workspace

blocked_commands:

- rm -rf

- curl

```

Key Commands

```bash

Interactive coding session

nemoclaw code --project /workspace/my-app

Code review mode

nemoclaw review --diff HEAD~1

Generate CUDA kernels

nemoclaw cuda --input matrix_multiply.py --optimize

Fine-tune on your codebase

nemoclaw customize --data /workspace/my-repo --method lora --epochs 3

```

NemoClaw vs OpenClaw vs Claude Code

| Feature | NemoClaw | OpenClaw | Claude Code |

|---------|----------|----------|-------------|

| GPU Required | Yes | No | No |

| Offline Mode | Full | With local models | No |

| Fine-tuning | Built-in (NeMo) | No | No |

| CUDA Generation | Optimized | Basic | Basic |

| Enterprise SSO | Yes | Plugin | No |

| Cost | AI Enterprise license | Free | $20/mo |