AI Agents Meet Modern SimOps-Compliant Cluster Administration

By Daniel Gruber, HPC Gridware, August 4, 2026

Most questions an HPC admin gets asked have the same shape: why is my job pending, which hosts have free GPUs, what changed in the config since last week. The answers live in the scheduler, but getting them out means qstat, qconf, and a fair amount of experience reading the output.

We built a different way in. Qontrol, our administration console for Gridware Cluster Scheduler (GCS), includes an MCP server. MCP is an open protocol invented by Anthropic that AI assistants use to connect to external systems. Point Claude or any MCP-capable assistant at it, and the assistant can read the live cluster: configuration, queue states, jobs, hosts, accounting data, daemon logs. Ask "why is job 4711 still pending" and it checks the actual scheduler state and tells you. Ask a new admin's question like "walk me through this cluster's setup" and it answers from the real configuration of your installation, not from generic Grid Engine knowledge in its training data.

Since GCS is part of the SimOps Software Stack, this fits a larger picture. SimOps-compliance means the software follows the community's documented best practices for automating and streamlining simulation operations - less manual scheduler archaeology, more time for actual engineering. SimOps 2.0 extends the framework toward agentic AI, and a scheduler that AI assistants can query directly is a concrete piece of that.

A note for anyone whose cluster data is sensitive, which is most of you: the MCP server is read-only, requires authentication, and the admin chooses which tool groups are exposed. Qontrol itself never talks to any AI vendor. You can run it against a cloud assistant or a local model in an airgapped network.

One thing we want to be upfront about: Qontrol and its MCP server are part of the commercial Gridware Cluster Scheduler product, not of the open source base. That base is Open Cluster Scheduler (OCS), the successor of Grid Engine, which we develop in the open - free packages are available from hpc-gridware.com, and the sources are on GitHub. If you run OCS today, you already have the same scheduler core that GCS builds on; moving to GCS adds the supported distribution and, with Qontrol, the web console and AI-agent integration described here.

The whitepaper covers the rest of Qontrol - verified live configuration changes, audit trail, guided workflows - and the MCP architecture in more detail.

→ Download the Whitepaper (https://whitepaper.clusterscheduler.io/)

→ Download Open Cluster Scheduler (https://hpc-gridware.com/download-main/)