Runcrate introduces an MCP server at https://mcp.runcrate.ai/mcp to automate GPU instance management for coding agents, addressing cost overruns and manual oversight. The server provides tools for listing GPU types with live hourly rates (e.g., H100 at $3.55/hr, RTX 4090 at $0.52/hr), creating instances with configurable max_runtime_minutes (5 min–7 days) and idle_shutdown_minutes (5 min–24 hrs), executing SSH commands, attaching storage volumes, and managing billing via a prepaid balance capped by workspace-defined monthly budgets. Agents interact via Streamable HTTP, requiring OAuth sign-in but no API keys. Integration with Claude Code, Codex, and Cursor uses simple CLI commands or one-click installs. Guardrails prevent agents from exceeding budgets or altering environment settings. Example workflows include running a training script on an A4000 GPU with a 2-hour max runtime and 30-minute idle timeout, or provisioning an RTX 4090 instance with vllm for Qwen inference, exposing port 8000 via HTTPS. Jobs automatically terminate after completion, ensuring no lingering costs. The system also offers 200+ open models via an OpenAI-compatible API (e.g., DeepSeek-V3) for non-GPU tasks. Pricing and setup details are available at runcrate.ai.