Claude Code Plugin

The sparkrun plugin for Claude Code teaches Claude how to manage LLM inference workloads on your DGX Spark systems.

What it does

• Slash commands — Quick actions for running, stopping, and managing inference jobs
• Skills — Detailed reference that Claude uses automatically when working with sparkrun

Installation

From the marketplace

# Add the marketplace (one-time setup)
claude plugin marketplace add spark-arena/sparkrun

# Install the plugin
claude plugin install sparkrun@sparkrun

Prerequisites

sparkrun CLI

# Install via uvx (recommended)
uvx sparkrun setup install

DGX Spark cluster

You need SSH access to one or more DGX Spark systems:

sparkrun cluster create mylab --hosts 192.168.11.13,192.168.11.14 -d "My DGX Spark lab"
sparkrun cluster set-default mylab
sparkrun setup ssh --cluster mylab

Claude Code

And you’ll obviously need to be using Claude Code.

Natural language usage

Just describe what you want — Claude uses the skills automatically:

• “Run the Qwen3 1.7B model on my cluster”
• “What inference jobs are running?”
• “Stop the nemotron model”
• “Show me available recipes for llama models”
• “Set up sparkrun on my DGX Spark cluster”

Key concepts

• Recipes are YAML files describing an inference workload (model, runtime, container, defaults)
• Runtimes are inference engines: vLLM, SGLang, llama.cpp
• Clusters are named groups of DGX Spark hosts
• Each DGX Spark has 1 GPU, so --tp N (tensor parallelism) = N hosts
• sparkrun launches detached containers — Ctrl+C detaches from logs, never kills the job