Claude Code can initiate a model‑training run on a Nodus cloud GPU with a single prompt, provided the user supplies five explicit inputs: the exact base‑model revision (e.g., a 40‑character Hugging Face commit like hf://Qwen/Qwen3-0.6B@c1899de289a04d12100db370d81485cdf75e47ca), the data source and its column mapping (a Volume named support‑chats containing chats.jsonl where the “question” column is the prompt and the “answer” column is the completion), a stop condition (maximum step count plus a validation split for early stopping), a cost ceiling (maxCostUSD field covering the job, retries, and export), and the desired output files (e.g., results.json and the adapter folder downloaded to ./runs/support-sft/). After installing the Nodus CLI, signing in, and attaching the local MCP server (pip install nodus-compute; nodus login; nodus mcp install claude), the user creates a Volume, uploads the dataset, and then sends the prompt to Claude Code. The agent constructs a TrainingJob manifest, returns a dry‑run estimate (for the example, ~40 minutes on an NVIDIA L4 priced at $0.43/hr → ≈$0.29 GPU time), awaits user confirmation, and then executes the job, polling status every ≤60 seconds. Upon completion, the adapter weights, metrics, provenance, and file checksums are saved locally; if the run suspends due to the cost cap, the user can raise maxCostUSD to resume from the latest checkpoint.