Apache 2.0 · Free & Open Source

Your Kubernetes cluster
has an AI debugger now.

KubeAstra investigates broken pods, finds the root cause with AI, and recovers your cluster. 51 tools. Auto-triage from Alertmanager. Runbook cache with citations. One agent. Zero guesswork.

View on GitHub Watch Demo
kubectl prometheus logs events topology
kubeastra investigate
$ kubeastra investigate --pod api-gateway --ns production

kubectl Collecting pod status and events...
metrics Pulling container resource usage...
logs Reading current + previous container logs...
events Correlating cluster events timeline...
analyze Running AI root-cause analysis...

Root cause: OOMKilled. Container RSS 142Mi exceeded limit 128Mi.
Trigger: Memory leak introduced in v2.3.1 (deployed 47 min ago).
Recommended: Rollback to v2.3.0 · blast radius: 2 services
Capabilities

Everything you need to debug K8s

51 built-in tools orchestrated by AI. From diagnosis to recovery, in seconds instead of minutes.

51 Investigation Tools

Pods, deployments, services, events, nodes, Helm releases, resource graphs, and more — plus GitOps repo search and Prometheus queries.

AI Root-Cause Analysis

Cross-references logs, events, and resource metrics to pinpoint exactly what went wrong and why. A synthesis critic verifies every claim against the raw evidence.

Automated Recovery

Pod restarts, deployment rollbacks, HPA scaling, YAML patches. Human-in-the-loop approval by default, or opt into automated recovery once you trust the plan.

Web UI Dashboard

Next.js interface with a live Investigation Trail, cost breakdown, and session sidebar. Watch the agent work, review findings, and approve actions.

MCP Server for IDEs

Model Context Protocol server exposes all 51 tools to AI-powered editors. Debug clusters from Cursor or VS Code without leaving the editor.

Fully Self-Hosted

Deploy inside your cluster with the bundled Helm chart. Your data never leaves your infrastructure. No SaaS, no callbacks, no telemetry.

Alert-Driven Auto-Investigation

Point Alertmanager at KubeAstra's webhook. Each firing alert triggers a scoped investigation and posts the root cause to Slack — before you open the laptop.

RAG Runbook Cache

Cached / grounded / cold retrieval router in front of the LLM. Sub-second answers on known failure modes, with citations to the source docs. Backed by Qdrant.

Multi-Step Remediation Plans

Complex fixes get proposed as atomic multi-step plans. Every destructive step exposes its dry-run output before executing, gated by a single-use confirmation token.

Tool Result Summarization

10,000-line kubectl logs get compressed to 3-line summaries before entering the LLM context. Heuristic first, LLM polish optional. Raw evidence stays one click away.

4 LLM Providers

Google Gemini, Anthropic Claude, OpenAI GPT, or Ollama (fully local). Swap providers with a single env var. Any OpenAI-compatible endpoint works — Azure OpenAI, vLLM, LiteLLM.

Workflow

How KubeAstra works

From broken pod to root cause in seconds, not minutes.

01

Detect

Point KubeAstra at a failing pod, or let it watch your cluster. It picks up CrashLoopBackOff, OOMKilled, ImagePullBackOff, pending pods, and other failure states.

02

Investigate

The agent runs kubectl tools automatically: pod describe, logs (current and previous), events, node status, resource usage, deployment history, network policies, PVC mounts, and more.

03

Analyze

AI processes all collected data together, cross-referencing signals across multiple resources to identify the root cause with confidence levels and supporting evidence.

04

Recover

Based on the analysis, KubeAstra recommends and can execute recovery actions: pod restarts, deployment rollbacks, or HPA scaling. Human approval required by default.

See It In Action

Demo

Watch KubeAstra investigate a failing pod, find the root cause, and suggest recovery.

Click to play demo
Built With

Tech Stack

Python
FastAPI
Next.js
TypeScript
Google Gemini
Kubernetes
Helm
MCP Server

Stop debugging the same way every time.

KubeAstra is free, open source, and ready to deploy. Star the repo, try it out, or contribute.