Skip to content

Plugin Commands

CommandDescription
/sparkrun:run <recipe>Launch an inference workload
/sparkrun:stop <recipe>Stop a running workload
/sparkrun:statusCheck status of running workloads
/sparkrun:list [query]Browse and search available recipes
/sparkrun:benchmark <recipe>Measure performance and resume saved work
/sparkrun:monitorLive host/workload monitoring
/sparkrun:proxy <action>Manage the selected inference gateway
/sparkrun:setupGuided setup for sparkrun and cluster config

Launch an inference workload on DGX Spark using a recipe.

/sparkrun:run qwen3-1.7b-vllm
/sparkrun:run nemotron3-nano-30b-nvfp4-vllm --tp 2
/sparkrun:run qwen3-1.7b-llama-cpp

Options: --hosts, --cluster, --solo, --tp, --port, --gpu-mem, -o KEY=VALUE, --dry-run

Stop a running inference workload.

/sparkrun:stop qwen3-1.7b-vllm
/sparkrun:stop nemotron3-nano-30b-nvfp4-vllm --tp 2

The --hosts/--cluster/--tp flags must match the original run command.

Check running workloads on your cluster.

/sparkrun:status
/sparkrun:status --cluster mylab

Browse and search available recipes.

/sparkrun:list
/sparkrun:list qwen3

Guided setup that walks through:

  1. Installing sparkrun
  2. Configuring a cluster
  3. Setting up SSH mesh
  4. Verifying the configuration
/sparkrun:benchmark qwen3-1.7b-sglang --cluster mylab
/sparkrun:monitor --cluster mylab --simple
/sparkrun:proxy status

Benchmarks wait for readiness and can resume saved work or retry publication. Use plain/JSON monitoring in noninteractive sessions. Proxy behavior follows the selected gateway; LiteLLM is the default on every channel, and SparkRoute requires separate installation and explicit selection. See Benchmarking and Proxy Gateway. A tune slash command is not enabled in the pinned companion plugin.

Claude also activates these skills automatically based on context:

SkillActivates when
runRunning, monitoring, or stopping inference workloads
setupInstalling sparkrun, configuring clusters, SSH setup
registryManaging recipe registries, creating/editing recipes