v1.1 · Kubernetes · AWS · GCP · Azure · Terraform

Turn Claude Code into
a DevOps engineer

42 specialist agents for Kubernetes, AWS, GCP, Azure and Terraform — incident triage, IAM review, plan review, CI/CD debugging, SLOs and postmortems. Every answer cites its evidence. Fixes ship as a pull request, never a live change.

npx github:shahneil76/opskit init

Works on any repo — Node.js 18+ and python3. Connect a Kubernetes cluster for live investigation, or use it against your manifests, Terraform and cloud configs without one.

Not sure where to start? Type one of these.

/k8s-triage

payment-service is in CrashLoopBackOff

Exit code, previous logs, recent rollouts — with the evidence cited.

/tf-plan-review

what does this plan actually destroy?

Blast radius, replacements, hardcoded secrets, cost outliers. No cluster needed.

/aws-iam-review

is this role over-permissioned?

Wildcard actions, stale cross-account trust, unused 90-day permissions.

/github-debug

why is this pipeline failing?

Root-causes a failing GitHub Actions run and gives the exact fix.

See OpsKit install itself

One command. It checks your cluster, wires the read-only connector, installs the safety hooks, and tells you which capabilities your specific cluster supports.

zsh — opskit init

$ npx github:shahneil76/opskit init

  ✓ Node.js v20.11.1
  ✓ context: production-eks
  ✓ Kubernetes 1.36 (v1.36.0)
  ✓ 34 skills, 42 agents, 47 commands
  ✓ safety layer installed (opskit-guard, MIT)
  ✓ 15 read-only tools ready

Active on this cluster:
  ✓ Sidecar-aware pod triage
  ✓ GitOps ownership resolution

Ready. Try: /k8s-triage <pod> is failing
Then:  npx opskit evidence bundle → attach the replay to the PR

0
Specialist Agents: 0
Investigation Skills: 0
Slash Commands: 0
Read-only MCP Tools

Generic AI guesses. OpsKit reads.

Four things go wrong when an AI agent meets real infrastructure. OpsKit was built around all four.

Your agent is guessing

Without a live connection it answers from training data. Fine for code. Not fine for an outage that started 11 minutes ago.

A read-only window into the real cluster — actual pod events, actual logs, actual resource state. Every conclusion cites the evidence it came from.

You can't trust it near prod

A confident wrong answer that runs kubectl apply is worse than no answer. Telling a model not to mutate is a prompt, not a control.

An allowlist hook blocks mutations at the OS level before they reach the cluster. It's MIT-licensed — read every line before you trust it.

Sidecar failures stay invisible

Since Kubernetes 1.36 sidecars report into initContainerStatuses. A pod shows Running while a sidecar CrashLoops, and naive triage calls it healthy.

k8s_get_pod_health normalises both arrays and tags every container as app, sidecar, or init. We proved the trap on a live cluster before shipping.

You can't prove what it did

The only record of an agent's actions is normally its own self-reported transcript. If a security team asks how you know it didn't do anything else, "trust me" isn't an answer.

Every tool call is appended to a hash-chained log. opskit replay --verify recomputes the chain and names the exact record if anything was ever altered, removed, or reordered.

Install in one command

Run npx github:shahneil76/opskit init in your project. OpsKit asks which cluster context to use, builds the MCP server, fetches the safety hooks, and verifies connectivity.

Node.js 18+, python3, kubectl. Claude Code, Cursor, Kiro, Codex CLI.

Give it your context

Fill in CLAUDE.md with cluster names, key services, runbook locations, and escalation paths. It's the context that turns a generic agent into one that knows your platform by name.

EKS, GKE, AKS, k3s, minikube, kind. Kubernetes 1.34–1.36.

Investigate, then fix

Type /k8s-triage payment-service is failing. OpsKit correlates events and logs, forms a hypothesis with cited evidence, and either opens a GitOps PR or tells you exactly what's missing.

47 slash commands · 42 specialist agents · all read-only.

Everything you need to go from alert to evidence

Eight capabilities that turn a general-purpose agent into one your on-call rotation can actually rely on.

MCP

Live cluster connection

A 15-tool read-only MCP server. Your agent reads real pod events, logs, and resource states — not documentation.

Diagnosis

Evidence contracts

Each of the 34 skills declares the evidence required before a conclusion. Missing evidence produces an abstention, not a guess.

K8s 1.36

Sidecar-aware triage

A sidecar CrashLooping inside a Running pod is invisible to anything reading only containerStatuses. OpsKit checks both arrays.

Safety

Enforced mutation block

An allowlist hook blocks kubectl apply, helm upgrade, terraform apply, and 120+ bypass patterns. Open source, auditable.

Least privilege

Scoped reads for pilots

Limit the agent to specific namespaces and label selectors so security can approve a payments-only pilot without handing over the cluster.

GitOps

Fixes arrive as PRs

Ownership resolved from ArgoCD or Flux annotations, a minimal patch, a rollback plan, and a verification checklist. Never a direct apply.

Trust

PR investigation replay

Redacted MCP output is captured during triage and bundled into collapsible PR comments, so reviewers replay exactly what the agent saw.

Multi-cloud

Cloud, Terraform, and beyond

42 agents across AWS, GCP, and Azure — Lambda errors, IAM wildcards, GKE drain storms, AKS quota limits, and plan blast radius.

42 agents. Not a list of names.

Every agent has one job and doesn't stray from it. These are the actual investigation protocols.

Kubernetes investigation

CrashLoop — exit codes, previous logs, config changes
OOMKill — limit vs. leak discrimination
Pending pods — quotas, taints, DRA claims
Network/DNS — NetworkPolicy, endpoints, CoreDNS
Rollout regression — pre/post deploy comparison
Sidecar failures — initContainerStatuses (GA in 1.36)
GitOps drift — ArgoCD/Flux ownership resolution
Scoped reads — namespace allowlist + label selector

Cloud investigation

AWS Lambda — timeout, OOM, throttles, DLQ analysis
AWS ALB 5xx — target health, access logs, WAF rules
AWS RDS — connection exhaustion, slow queries, failover
EKS — Karpenter, aws-auth, IRSA, OIDC, add-on conflicts
GCP — GKE drain storms, Cloud Run cold starts, Cloud SQL
Azure — AKS vCPU quotas, App Service throttling, SQL DTU
IAM/RBAC — wildcards, Editor roles, subscription-scope Contributor
Cost — idle NAT Gateways, BigQuery slots, committed-use gaps

Terraform / IaC

Plan review — blast radius, destroys, replacements
Security scanner — public S3, open SGs, unencrypted RDS
Secret detection — credentials in source and plan output
Cost estimation — expensive types, storage anomalies
State drift — orphaned resources, import candidates

Platform and SRE

Incident commander — timeline and stakeholder updates
SLO designer — multi-window burn-rate alerts
Postmortem analyst — prevention tasks with acceptance criteria
Capacity planner — usage trends, growth, headroom
Deployment guard — missing probes, :latest tags, limits

299/299

Tests passing

129 safety bypass · 64 K8s live cluster · 45 AWS · 40 Terraform · 20 flight recorder tamper

100%

Mutation block rate

0 bypasses across 129 adversarial cases — flag injection, aliases, pipe chains

MIT

Safety layer licence

Every line of the hook engine is public at github.com/shahneil76/opskit-guard

< 15 min

To first investigation

One command from install to a working cluster connection

Free · no signup

Kubernetes YAML beautifier & linter

Paste a manifest. Get it formatted, validated, and checked against the same production rules OpsKit's agents apply during an investigation. Everything runs in your browser — nothing is uploaded.

Open the tool All free tools

YAML beautifier

Normalise indentation, key order, and quoting across every document in a multi-doc manifest.

Manifest linter

16 production checks — missing limits, :latest tags, absent probes, privileged containers, plaintext secrets.

YAML ↔ JSON

Convert either direction for kubectl --dry-run output, Helm values, or CI pipeline configs.

Free. MIT licensed. No catch.

No subscription, no seats, no usage metering, no purchase at all — the whole kit, including the safety layer.

Open source kit

MIT

Freeforever · open source

For Staff SREs and Platform Engineers who want their AI to actually read the cluster.

42 specialist agents (Kubernetes, AWS, GCP, Azure, Terraform, SRE)
34 investigation skills with evidence contracts
47 slash commands
15-tool read-only MCP server
Claude Code, Cursor, Codex CLI, Kiro & other MCP agents
MIT licensed — fork it, audit it, ship it
opskit-guard safety layer (MIT, open source)
Flight recorder — tamper-evident, hash-chained action log

Get OpsKit — Free

No signup

install and go

Open source

MIT licensed

No lock-in

fork it anytime

Delivered

via public GitHub repo

If it saves you an incident — buy me a coffee

Frequently asked questions

Everything worth knowing before you point an AI agent at a production cluster.

What exactly gets installed?+

OpsKit drops files into your project's .claude/ directory: 42 agent files, 34 skill files, 47 command files, a configured MCP server, and the safety hooks. The MCP server connects to your cluster using your existing kubectl credentials. Nothing is installed globally.

Which AI tools are supported?+

Claude Code, Cursor, Codex CLI, Kiro, and other MCP-capable agents can use OpsKit's read-only cluster tools. Claude Code and Kiro also get full safety hook enforcement at the OS level. Cursor and Codex CLI connect for investigation; mutation blocking depends on their hook model, so for strict production guardrails we recommend Claude Code or Kiro.

What are the requirements?+

Node.js 18+, python3, and kubectl. That's it. OpsKit uses your existing kubeconfig — it never asks for credentials beyond what kubectl already has. Kubernetes 1.34, 1.35, and 1.36 are tested across EKS, GKE, AKS, k3s, minikube, and kind.

How do I install it?+

OpsKit is free and open source (MIT) on GitHub. Run the one-line install command from the README — `npx github:shahneil76/opskit init` — and it's working in minutes. No purchase, no license key, no zip files.

What stops it from breaking things?+

A pre-tool hook intercepts every bash command before it executes and checks it against an allowlist. kubectl apply, helm upgrade, terraform apply, kubectl exec, and 120+ bypass patterns are blocked at the OS level before reaching your cluster. The safety layer is open source — you can read exactly what it allows and blocks before trusting it with production.

Is it really free? What's the catch?+

No catch. OpsKit and its safety layer (opskit-guard) are both MIT licensed — fork it, audit it, modify it, ship it. If it saves you an incident, there's a Buy Me a Coffee link in the README, entirely optional.

Still got questions? support@runopskit.com

Your next incident deserves evidence.

Stop pasting logs into a chat window. Give the agent the cluster — read-only.

Get OpsKit — Free Try the free YAML tool

Turn Claude Code intoa DevOps engineer