Ventrolab

Products / AI SRE

Open core Apache-2.0 · v0.4.0

The on-call engineer that reads the alert before you do.

Ventrolab AI SRE takes an alert from Alertmanager, investigates it across your clusters and hosts, comes back with a root cause and, when a playbook fits, a fix — executed only within the limits you set.

Try the demo View on GitHub
AI SRE web console
Ventrolab AI SRE web console: open incidents, time to diagnosis, component health and latest incidents with root-cause confidence

01 How it works

  1. Ingest. Alertmanager, a scheduled check or a chat message sends an event. Alerts about the same object are correlated into one incident.
  2. Investigate. HolmesGPT picks matching skills from the knowledge base and reads metrics, logs, events and specs with read-only tools.
  3. Explain. The result is a structured verdict: root cause, confidence and the playbook that fits, if any.
  4. Decide. Action Guard applies the mode, risk class and limits: propose only, ask for approval, or execute.
  5. Act and verify. Prechecks, dry run, action, PromQL verification and rollback. Every step lands in the audit log and in chat.

02 What you get

Root cause, not noise

Incidents with timeline, alerts and an RCA with confidence. Route only the alerts that need a cause found; keep direct receivers for the rest.

Guarded remediation

SRE_MODE readonly or autofix, risk classes per playbook, approvals and rate limits. Write access is opt-in per agent.

Every cluster and host

k8s-agent for remote clusters and host-agent for VMs, over an outbound tunnel with tokens or mTLS. Secrets are redacted before anything reaches the LLM.

Your LLM, your data

Claude, OpenAI, vLLM or Ollama through LiteLLM, with fallbacks. Prometheus, Loki, VictoriaLogs and Elasticsearch/OpenSearch as data sources.

Console, chat, CLI

Dashboard, approvals, ad-hoc questions, playbook catalog and audit log. Roles viewer, operator, admin; local accounts or OIDC.

Knowledge you can extend

Troubleshooting skills and remediation playbooks are plain files: add your own through Helm values or contribute them upstream.

03 Architecture

One control plane per installation, agents wherever your workloads run. Everything ships as a Helm chart.

Architecture: Alertmanager and chat feed Event Ingest and NATS; SRE Controller drives HolmesGPT through the LLM Gateway; Action Guard and Tool Hub reach local and remote clusters and hosts

04 Quick start

A local kind cluster with monitoring, AI SRE and three broken workloads, investigated within minutes. Needs Docker (8 GB RAM), kind, helm, kubectl and an Anthropic or OpenAI key.

# clone and run the demo
git clone https://github.com/ventrolab/ai-sre.git && cd ai-sre
export ANTHROPIC_API_KEY=...   # or OPENAI_API_KEY=... LLM_PROVIDER=openai
make demo                      # prints the console URL and the admin password
make -C deploy/demo status     # incidents and their root causes

Production: Helm chart oci://ghcr.io/ventrolab/ai-sre/charts/ai-sre.

05 Editions

Community

Free · Apache-2.0

  • Full investigation pipeline and web console
  • Autofix for low-risk playbooks
  • Approvals in the web console
  • Up to 5 agents
  • Local accounts and OIDC sign-in
Get it on GitHub

Enterprise

Signed license, works offline

  • Everything in Community
  • Autofix up to medium risk; high always needs approval
  • Approve and Reject right from chat
  • Unlimited agents
  • Audit log export (JSON, CSV)
  • Signed multi-arch images with SBOM and provenance
Contact us