AWS open-sourced a HyperPod control plane that exposes 38 MCP tools with parameter validation before execution

AWS Machine Learning Blog·medium signalXBlueskyLinkedInCopy link

HyperPod InstantStart (github.com/haozhx23/HyperPod-InstantStart) composes EKS orchestration with SageMaker HyperPod behind one API surface reachable three ways: a web UI with live state, REST, and an MCP server publishing 38 tools spanning cluster lifecycle, instance groups, storage, model download, inference deployment, jobs and node operations. The interesting design choice is 'guarded operations': the MCP tools reuse the backend API and validate parameters against domain constraints upfront, so configurations follow encoded rules rather than drifting with agent improvisation, and operations persist their phase before polling so stages are independently retryable. EKS control-plane creation runs roughly 8-12 minutes.