Built by someone who was on call

The problem it was built for

A general-purpose AI agent is a historian. It has read every Kubernetes doc ever written and understands the architecture in the abstract — and then it guesses about your cluster, confidently, because it cannot see it. That's fine when you're writing code. It is not fine during an outage that started eleven minutes ago.

The tell is when the agent asks you a question your cluster could have answered instantly. "What were the previous container's logs?" "Is this a resource constraint?" You copy, you paste, you wait. The most capable tool on your desk is somehow less useful than a colleague who knows where to look.

What we actually built

A 15-tool read-only connection to the live cluster, 34 investigation skills that declare what evidence they need before drawing a conclusion, 42 specialist agents, and a safety layer that blocks every mutation verb before it executes.

The safety layer came second, and it came from fear. Giving an AI agent a live cluster connection means giving it a loaded gun. So the block is an allowlist rather than a denylist, it's enforced at the OS level rather than requested in a prompt, and it's MIT-licensed and public. If we're asking you to point an agent at production, you should be able to read exactly what's standing between it and your cluster.

What we won't claim

This matters more to us than the feature list, because the failure mode of AI ops tooling is confident nonsense.

• It will not replace your SRE team. It's a faster way to gather and cite evidence. The judgement is still yours.
• It does not apply fixes. The output of a remediation workflow is a patch or a pull request. Merging is your call, through your normal review path.
• It abstains rather than guesses. When a skill runs out of the evidence its contract requires, it says so. An investigation that admits a gap is worth more than one that fills the gap with plausible reasoning.
• We publish what isn't verified. Our comparison pages say where the alternatives are better, and three of the four are free and open source. The test breakdown states which suites run against a live cluster and which cloud packs have no adequate local emulator and are therefore unverified content.

How it's sold

One-time purchase, $19 at launch pricing, single-user perpetual licence with 12 months of updates. No subscription, no seats, no usage metering. Delivered as access to a private GitHub repository. Full terms are on the Terms of Service page, including the refund policy — all sales are final, and we'd rather you knew that before you clicked than after.

The free tools are genuinely free and require no signup, because the cheapest way to judge whether we know what we're doing is to use something we built.

Who's behind it

OpsKit is built and maintained by [LEGAL ENTITY NAME] — a working DevOps engineer, not a venture- funded team. That's the honest scale of it. It also means the person answering your support email is the person who wrote the code.

Email support@runopskit.com. Bugs and feature requests are better as issues on GitHub.