The range is the unusual part: the same server that scales a deployment or submits an sbatch job also answers which agent spent the most last month, verifies the audit chain cryptographically, and tracks AI Act controls. That is only workable because it runs against your own VibOps instance through a gateway inside your infrastructure rather than reaching clusters directly — and because every write is recorded, which is what makes letting an agent deploy and scale defensible in the first place.
A client for a VibOps instance, which sits between your assistant and a GPU fleet. Every operation runs through your VibOps deployment by way of a gateway installed in your own infrastructure, and lands in an immutable audit log.
- Clusters listed with GPU utilisation, plus live Kubernetes deployment status with replica counts, pod health and resource usage
- Deployments scaled, models deployed onto a cluster, Helm charts installed and uninstalled, and kubectl run directly where no dedicated tool covers the operation
- Repositories cloned onto the gateway — with Kubernetes manifests applied automatically when a cluster is named — and automation pipelines triggered
- Slurm: cluster info with partitions, node states and GPU resources, jobs listed and their status read, stdout tailed to follow training, multi-node GPU jobs submitted through a generated sbatch script with a dry run available, and jobs cancelled
- Virtual machines across Proxmox, Xen Orchestra and vSphere — listed, started, stopped, live-migrated and snapshotted, with ESXi hosts and real-time VM metrics
- Container registries — Harbor, ECR and Google Artifact Registry — repositories and tags listed, an image checked for existence, and a stale tag deleted behind an explicit confirmation
- Observability: hourly GPU utilisation series, workload breakdown by type, MTTR on GPU alerts, job throughput with p50 and p95 latency, and cost estimates once a per-cluster rate exists
- Anomalies detected automatically across the fleet — idle GPUs, spikes, hardware errors — listed, filtered to those still open, and resolved by hand when the cause was fixed elsewhere
- FinOps: spend trend by cluster and tenant, monthly chargeback, VM cost attribution, waste analysis for idle resources, and per-agent LLM usage with monthly budgets enforced at the proxy
- Governance: the immutable audit log queried and its HMAC-SHA256 chain verified, organisation policy read and replaced, SOC 2, RGPD and HIPAA reports generated asynchronously, and AI Act controls tracked with a compliance score
- Agent machine identities created, rotated and revoked — the raw key returned once only — model access rules by glob pattern, and a dependency graph of which agents call which models, connectors and sub-agents
- LLM-as-judge evaluation of a completed job against a rubric, scored per criterion
- Gateways registered and revoked, secrets stored encrypted with values never returned, and LDAP and SIEM configuration — Splunk or Datadog — read and updated
A VibOps instance, with a gateway installed inside the infrastructure you want it to reach. Python 3.11 or newer, installed from the repository with pip. Two environment variables: VIBOPS_URL for your instance and VIBOPS_TOKEN for an API token created under Settings → API Tokens. Cost estimates return nothing until a per-cluster rate is set, and setting one needs the organisation admin role. Every write tool records what it did in the audit log, and deleting a registry tag requires an explicit confirmation flag.
One command plus a key — pip install git+https://github.com/VibOpsai/vibops-mcp.git, then supply credentials