Your clusters. Your data. Your call.
Every other AI-SRE tool wants your telemetry in their cloud. oneinfra is the opposite: the entire platform — agents, control plane, storage, AI — runs inside your network. The only thing that ever leaves is what you wire to leave.
Your data never leaves
Agent, control plane, AI, storage — all of it runs inside your VPC. Cluster telemetry, logs, traces, deploy history, cost data, investigation transcripts: never copied off-network. The only thing that ever talks to the outside world is your own LLM provider — and that's wired to your account, not ours.
Bring your own model
OpenAI, Anthropic, AWS Bedrock, Azure OpenAI, or a self-hosted model (Ollama, vLLM, anything OpenAI-compatible). Switch with one config line. Your prompts and context never train third-party models. Air-gapped operation supported end to end.
No lock-in, ever
Source-available. Helm-installable. Investigations, evidence, audit log all live in your own SQLite + object store. Walk away anytime, take your data with you. The closed-source SaaS competitors' churn rate is your retention strategy.
Extensible by design
MCP (Model Context Protocol) is a first-class citizen. Wire any tool — internal runbook, custom dashboard, proprietary database — without forking. The agents call it, the eval harness measures it. Most teams ship a custom adapter in under an hour.
One Helm command. No code changes.
Deploy into any Kubernetes cluster — EKS, GKE, AKS, k3s, kind, OpenShift, your own. Read-only by default. Promote to writer with explicit RBAC when you trust it.
# Add the chart
helm repo add oneinfra https://charts.oneinfra.tech
# Install with your model provider of choice
helm install oneinfra oneinfra/oneinfra \
--namespace oneinfra --create-namespace \
--set llm.provider=anthropic \
--set llm.apiKeySecret=anthropic-key
# That's it. Open the UI:
kubectl port-forward -n oneinfra svc/oneinfra 8080:80Note: charts.oneinfra.tech is the public chart repo (coming with GA repo open-sourcing — until then, install via the Helm chart in the source repo).
Things people ask before going self-hosted.
Doesn't self-hosted mean we run our own LLM too?
No — only if you want to. The default is "your VPC calls OpenAI / Anthropic / Bedrock with your API key, we never touch it". Air-gapped is the option for regulated environments.
Won't self-hosted be slower than SaaS?
Latency-wise: the AI call dominates everything. Same model = same speed. The only difference is the telemetry stays close to the cluster instead of round-tripping through a SaaS region.
How is this different from k8sgpt?
k8sgpt is a CLI that runs ad-hoc analyses. oneinfra is the full agentic platform around it — long-running watchers, multi-cluster federation, cost FinOps, SLOs, Slack incident threads, evidence vault, and a UI. We even integrate k8sgpt as one of our analyzers.
Who maintains it if it's open source?
A small commercial team funds development; the source is open so you can fork, audit, or self-support if you ever lose faith in us. That's the whole point of the model — we earn our keep by shipping faster than you'd need to fork.
No demo call required.
The live app is at oneinfra.tech/oneinfra — explore, then helm install when you're ready.