AI changed how code gets written in two years. Why did it barely reach the people who run it?
That question has followed me around for two years. I have spent most of my working life keeping other people's systems alive, and I watched developers get a new way to work while the people on call got one more dashboard.
Today, after 136 days of building in the open with more than 300 operators, Kepler opens to everyone. It is an IDE for operations: a desktop app where you and an AI work on live systems together. It learns your systems, shows its proof, and asks before it changes anything.
Think of Kepler as the repo production never had.
That line needs explaining, and so does everything else we claim. So here is the reasoning, and then the receipts.
Production has no Ctrl+Z
Coding agents look like magic for a reason almost nobody mentions. The model is standing on a repo, and a repo quietly does four things for it.
- It holds still. Everything the agent needs is in one folder, and it does not change while the agent reads it.
- It remembers. Git has every change, who made it and when.
- It checks the work. Tests fail and the compiler complains, in seconds, without a person.
- It forgives. The worst day is a revert.
Production does none of them. It moves while you look at it. It forgets the hotfix someone applied at 2 AM and never committed. It has no test for "is this the root cause", so a wrong answer looks exactly like a right one until a customer tells you which it was. And it has no Ctrl+Z. Restart the wrong pod during a flash sale and you have not broken a function. You have broken a transaction, and somebody's trust.

At SREday in Bangalore this year I asked a room of people who carry pagers whether they would let an AI restart a production database right now. The gut answer is no. It was mine too. That is not fear of new technology. It is an accurate memory of what it costs to act on a live system with no undo.
I have written about this before, in the wrong substrate and why AI for operations can't live in one repo. Connecting a coding agent to your cluster gives it access. It does not give it any of the four.
What Kepler is
The IDE is the part you work in. It runs on your machine, with your access, on macOS, Windows and Linux. It has a terminal, a browser and an editor that you and the agent share.
The harness is the part the AI works in. It builds, on purpose, the four things production will never hand over.
- A map, because production won't hold still. Kepler learns your systems: the services, the hosts, the conventions, how the parts connect. The map goes deeper with every incident, and watchers tell you when something moves.
- Memory, because production forgets. What you teach it stays. You can read what it believes and correct it. A procedure becomes memory only after it has run and worked.
- Proof, because production has no tests. Every finding comes with the commands it ran, what it saw, and what it is only guessing. You are the referee, and it hands you what you need to rule.
- Asking first, because production has no undo. It asks before it changes anything. You set what it may do, and it can never loosen its own limits.
The model is the one part we chose not to own. Bring your own Anthropic, OpenAI, Google or Bedrock key on any plan, or use ours. Models will change every few months. Your map and your rules should not reset when they do.
The receipts: what operators said
The first commit was on 7 July. Eighteen days later a few operators had an alpha. In September, 76 operators ran the beta on real systems and sent 7,567 requests in twelve days, 3,095 of them on the busiest day. They were not gentle. Here are four things they told us, word for word, and what we did.

"When something breaks I want an answer now, not an essay."
We had assumed depth was the product. Then we saw that 86% of requests went to the fast model. So fast became the default, and depth is one step away. Speed turned out to be the first form of trust.

"Too many approval walls. I can't work like this."
The lazy fix was fewer approvals, and it would have traded away the whole point. The real fix was memory. You approve once, Kepler keeps the rule, and it still asks for anything new. Asking first only works if it remembers the answer.
"My model keys are sitting in a readable file."
They were right. We said so, and fixed it the next day. Operators audit for a living. You do not earn them with a launch. You earn them by being checkable.
"It says it's reconnecting when nothing is wrong."
Fixed the next day. We rebuilt how the app talks to the agent, so a busy session no longer looks like a lost one. A tool that raises false alarms about itself will not be believed about your systems.
Most AI tools for operations want your production data in someone else's cloud. Kepler runs on my machine, works with the model I pick, and drives a real browser through my dashboards. It does the legwork and I never leave my flow.
Mridul Gain, SMTS, Platform9
On 25 October, 146 people spent a day building with Kepler in one room in Pune. More than 300 operators have shaped it so far, and they have written 168 public posts about what they built. This launch is theirs too.
What it can't do yet
Kepler ships every week and it is not finished.
- The Mac build is for Apple silicon. There is no Intel Mac build yet.
- It is built for one person. A team cannot sign up by itself yet. If you are on call as a team, talk to us.
If something breaks or feels wrong, tell me on Discord. I would rather hear it now than after it is polished.
Why we are betting on this
Writing software is no longer the bottleneck. Running it is. AI has made code cheap. Every team ships more of it, faster, and all of it lands in production, where the number of people who understand the whole system has not changed.
The operator has never been the customer. Monitoring is bought for the infrastructure. Paging is bought for the organisation. The developer got a tool bought per seat, to make one person better. The person on call got a notes file. That is why Kepler is a tool for the operator: personal, on your machine, free for one person.
I have seen what happens when the harness holds. A month ago I gave Kepler a job I had been putting off for months: move our CI/CD to GitOps. It opened a war room, sent subagents through every service and dependency, planned the migration and carried it out, checking each step. It messaged me on WhatsApp whenever it needed an approval. It was done in under five hours. I would have put that work at over a month. I was not braver that night. I had a harness.
Available today
Kepler is free on one machine, for good. That includes chat, the terminal, the browser and the editor, memory, MCP and the CLI, up to three watchers, and your own model keys.
Pro is $20 a month, or $16 a month billed yearly. It adds war rooms and topology, automations, remote access, unlimited watchers and hosted models, on three devices. Every new account starts with 14 days of Pro, no card.
Download Kepler, read the docs, and tell us what it got wrong.
Developers did not get braver. They got a place that forgives. Code has a harness. Now operations does too.



