A Human-in-the-Loop RCA Agent for Kubernetes — LLM Diagnosis, OPA-Bounded Remediation
📖 Overview
K8gentS is a Kubernetes diagnostic agent built around a single premise: an LLM is well-suited to about cluster failures, but unfit to on them unsupervised. It continuously watches the cluster event stream, runs -class events through a Gemini-powered pipeline, and surfaces the top hypotheses with resolution steps to Slack. Every remediation is — a reviewer approves or rejects via interactive buttons before anything mutates. Approved fixes execute inside an ephemeral Job whose blast radius is bounded by an admission policy at the API server, independent of RBAC. The reasoning is non-deterministic by design; the controls around it are not.
Kubernetes RCA agent that sandboxes its own AI with OPA Gatekeeper. MCP-ready for any client.
API key
AI_API_KEY, AI_MODEL, KUBECONFIG
पिछली जाँच
2026-11-09
स्टार
—
टूल
—
अपटाइम
—
reason
act
Warning
root cause analysis (RCA)
human-in-the-loop (HITL)
OPA Gatekeeper
The goal: reduce MTTR on diagnosable failures without handing an LLM a kubectl apply.
🧠 The Hard Problem
Building the Kubernetes side of this is straightforward. The real challenge is making a diagnostic layer trustworthy when the engine behind it is fundamentally non-deterministic.
Traditional observability is built on guarantees. Alerts fire on known thresholds. Dashboards show reproducible numbers. Logs return consistent answers to the same query. SRE success depends on that predictability — it's what makes incident response repeatable and on-call sustainable.
An LLM-driven diagnostic layer breaks that contract. The same pod failure can produce three different plausible explanations across three different runs. Each may be coherent. Each may even be correct under different assumptions. But "plausible" is not the same as "right," and for infrastructure, the gap between them is where outages live.
Some of the specific problems I've been working through while building K8gentS:
1. Confidence scoring with an unbounded output space.
The agent returns top-3 root causes with confidence metrics, but confidence in what, exactly? The model is not selecting from a fixed set of known failure modes — it's generating free-form hypotheses. A calibrated confidence score needs a reference distribution, and the distribution here is whatever the model happened to produce this run.
2. When to trust reasoning vs. fall back to deterministic checks. - THE ART
Some failures (CrashLoopBackOff, OOMKilled) have well-traveled diagnostic paths and a deterministic check will be right every time. Others benefit from the model's ability to interpolate across signals. Drawing that line — and doing it at runtime — is non-trivial and where the art really lies.
3. Evaluating an agent that's supposed to find failures you didn't anticipate.
The standard ML evaluation approach assumes you know what "correct" looks like. For a diagnostic agent, part of the value is catching novel failure modes — by definition, failures you couldn't pre-enumerate. So how do you decide what is wrong and what is right?
4. The "Tool in the Cluster" problem.
How do you monitor the monitor? Currently the service is set to run as a service in the cluster, but what if the service itself causes resource exhaustion, or is experiencing failures itself? How can you identify if the service itself is the cause of your issue?
5. Determining the right model for this problem.
With so many other types of Machine Learning models out there, is a non-deterministic large language model really the right choice or is another model better suited for infrastructure type problems?
These are the questions I'm actively working on. If you've solved any of them — or have a sharper framing than I've got — I'd like to hear it.
✨ Core Responsibilities
Continuous Monitoring: Watches the cluster for error events, crashed pods, CrashLoopBackOff states, OOMKilled events, and other failure conditions such as Connectivity/DNS, Database Deadlock, or Secret/Config Missing.
Automated Root Cause Analysis (RCA): Upon detecting an anomaly, it securely fetches relevant context (recent logs, pod descriptions, event history), sanitizes it of secrets/PII, and sends this context to an LLM-based reasoning engine.
Notification & Confidence Scoring: Notifies your designated Slack communication channel via Socket Mode with:
A descriptive summary of the error.
The top 3 possible root causes, each with an associated confidence metric.
Step-by-step resolution instructions.
Interactive Remediation (Opt-in): Prompts the user directly in Slack with interactive buttons: "Approve Fix" or "I'll do it manually".
Default Posture - Read Only: The agent is strictly READ-ONLY, making it exceptionally secure by default.
🛠️ Architecture & Security
The agent is designed so that each layer independently limits blast radius — not as redundancy for its own sake, but because no single control is sufficient when the reasoning engine is non-deterministic.
Layer
Mechanism
What it prevents
Pod security
runAsNonRoot, read-only filesystem, all Linux capabilities dropped
Container escape, privilege escalation
RBAC
Agent pod is strictly read-only; write verbs live only on k8gent-executor-sa
Agent compromise → cluster mutation
Ephemeral executor
Short-lived Jobs via k8gent-executor-sa; ttlSecondsAfterFinished=120
Persistent foothold after remediation
OPA Gatekeeper
Rego policy enforced at the API server admission layer
Executor escaping its scope, even if RBAC is misconfigured
Log sanitization
Regex sweeper strips IPs, JWTs, API keys, emails before LLM call
Secrets exfiltration via LLM prompt
Rate limiting
Hourly circuit breakers and event debouncing
Noise-driven API budget exhaustion
Ingress-free comms
Slack Socket Mode; no exposed endpoints or Ingress rules
Inbound attack surface
The OPA Gatekeeper policy (defined in deploy/helm/k8gents/templates/opa-gatekeeper/) explicitly blocks the executor service account from modifying serviceAccountName, enabling hostNetwork or hostPID, operating inside kube-system, or mutating any resource kind other than pods and deployments — enforced directly at the Kubernetes API admission layer, independent of RBAC.
🚀 Deployment
K8gentS ships as a Helm chart. OPA Gatekeeper is a declared chart dependency — the security sandbox installs automatically alongside the agent.
Prerequisites
Kubernetes v1.20+, Helm 3
kubectl authenticated to the target cluster
A Google Gemini API key (AI_API_KEY)
A Slack app with Socket Mode enabled (generates SLACK_BOT_TOKEN starting xoxb- and SLACK_APP_TOKEN starting xapp-)
Slack channel ID (SLACK_CHANNEL_ID) and a comma-separated list of approver Slack user IDs (ALLOWED_APPROVERS)
1. Configure Slack
Create a Slack App at api.slack.com.
Enable Socket Mode → generates an App-Level Token (xapp-...).
Enable Interactive Components.
Add chat:write and chat:write.public OAuth scopes → generates a Bot Token (xoxb-...).
Invite the bot to your alert channel and copy the Channel ID from channel settings.
Confidence calibration — the current scoring reflects the model's self-reported certainty, which doesn't reliably correlate with empirical accuracy.
Deterministic routing — canonically diagnosable failures (CrashLoopBackOff, OOMKilled) shouldn't route through the LLM at all. Building a reliable runtime classifier for "known answer" vs. "needs reasoning" is the next structural change.
Evaluation — regression testing an agent designed to catch novel failures requires a framework that doesn't yet fully exist for this problem domain. Synthetic failure injection (chaos engineering) is the most promising direction, but coverage is inherently limited.
Post-remediation verification — after executing a fix, monitor the target namespace for 60s and post a follow-up Slack thread confirming recovery or flagging that the crash state persists.
अभी टूल सूची नहीं है। रिमोट सर्वर को पहले हेल्थ चेक के बाद सूची मिलती है।
FAQ
सवाल
K8gentS — Kubernetes RCA MCP Server कैसे इंस्टॉल करें?
इस पेज पर Cursor, Claude Desktop, Claude Code या VS Code के लिए stdio कॉन्फ़िग कॉपी करें। यह पब्लिशर का इंस्टॉल इस्तेमाल करता है, PlugMCP प्रॉक्सी नहीं।
क्या K8gentS — Kubernetes RCA MCP Server के लिए API key चाहिए?
पब्लिशर ने ये एनवायरनमेंट वेरिएबल बताए हैं: AI_API_KEY, AI_MODEL, KUBECONFIG।
K8gentS — Kubernetes RCA MCP Server कौन सा ट्रांसपोर्ट इस्तेमाल करता है?
K8gentS — Kubernetes RCA MCP Server stdio पर लोकल चलता है।
K8gentS — Kubernetes RCA MCP Server की पिछली जाँच कब हुई थी?
पिछली जाँच 2026-11-09 को हुई। Kubernetes RCA agent that sandboxes its own AI with OPA Gatekeeper. MCP-ready for any client.