cloudgpu-mcp (CloudGPU MCP server)

cloudgpu-mcp is an MIT-licensed MCP server that lets MCP clients rent a per-minute GPU, run a job, download results and shut it down. It runs locally over stdio (npx -y cloudgpu-mcp) and uses the CloudGPU account behind your CLOUDGPU_API_KEY. An agent can pick a GPU (RTX 3090, RTX 4090, RTX 5090, A100, H800 or RTX PRO 6000, depending on the live deploy page), start it from a template such as Ollama, ComfyUI, JupyterLab or LLaMA-Factory or as a raw SSH box, run commands, pull back output files and destroy the instance when the work is finished. Billed per minute, rounded up, with a 1-minute minimum; billing stops when you destroy the instance. The server is also listed in the official MCP Registry as io.github.mrhy7176/cloudgpu-mcp. The same CloudGPU key works with an OpenAI-compatible model API at cloudgpu.app/v1.