Frequently asked questions

Yes. You say what you want, like "summarize this document" or "organize my Downloads folder", and Kernl does it.

The free version does the work: it reads, writes, edits and runs things in one project folder, with no time limit and no account or card. It asks you first before anything it cannot take back. $59 a year buys what you do with that: routines that run on a schedule or a trigger, a separate project folder for every job, answering a question from another device while a run waits, away from your desk, and deciding how far it goes on its own. There is a 90-day trial before that starts, and no card to begin it. Cancel any time and the version you have keeps working. Companies need a commercial license. Full details here.

If you are on Windows 10 or 11, almost certainly. A graphics card from the last few years makes the built-in AI fast. Without one it still works, just slowly, and you can skip the built-in AI altogether by connecting an account you already have with Claude, ChatGPT, Gemini, or any of the other providers Kernl can talk to. Windows today. macOS is in progress and Linux is planned. There is more detail under "Do I need a graphics card?" below.

No. Your conversations, file contents, and everything you do in Kernl stay on your computer. We don't collect them and can't read them.

It does things. Ask it to tidy your Downloads folder, or to fix a bug and run the tests, and it works through your files, your programs, and your browser, then checks whether the result is right. Anything risky stops and asks you first.

Yes, and that part is what the paid version is for. A routine is a job you set up once and then leave alone: it runs on a schedule, or whenever a file in a folder you choose changes, or whenever you press go. It works in a real session you can read afterwards, and if it hits something it is not allowed to do on its own, it stops and waits for you rather than guessing. You can also pair a second device so the waiting question reaches you away from the machine doing the work. What the licence covers.

Yes. A skill is a small pack that tells Kernl how one job is done: what it is for, the steps, what counts as finished, and what it is allowed to touch while doing it. Several come with the app for jobs like tidying a folder, going through a pile of documents, or writing up a review, and you can install one of your own from a folder on your machine. A skill narrows what Kernl may do rather than widening it, so turning one on cannot hand it more access than you had already given.

Go to the download page and download it. There is no account, no access code, and no sign-up.

The installer is about 252 MB. The first time you run it, if you use the built-in AI, it downloads one that suits your machine. That is about 1.8 GB on a laptop or a machine without a graphics card, and 19.7 GB for the largest that fits a single big card. There is a much larger one, 133.5 GB, for multi-GPU machines. It only downloads once. If you connect an account of your own instead, there is nothing extra to download.

No. Everything it needs comes in the one installer: a real browser it can drive, web search and page reading, document and picture tools, reading text aloud, and turning a recording into text. In all, 15 extra sets of tools are switched on the moment you open it, with nothing to sign up for. The accounts only you can connect, like GitHub or Slack, wait in Settings, Extensions, and stay off until you connect them. There are 54 of those.

The same idea, in that it does the work rather than only describing it, but for anything rather than only for writing code, and it runs on your own machine. Your files and your work stay there, and you choose whether it uses the built-in AI or an account of your own.

Kernl downloads the new version and installs it. Your conversations, settings, and files are not touched, and nothing else is sent anywhere.

It is early. Expect the occasional rough edge. The installer is signed, though Windows may still warn you the first time you run it, because that prompt lifts on how many people have downloaded a build rather than on the certificate. Windows today. macOS is in progress and Linux is planned.

The Kernl Discord gets you an answer soonest. Bugs, questions, and feature requests all go there, and new builds are announced there. You can also email hello@getkernl.ai. To make a report worth answering, open Settings, About, System and save a support bundle: it is a small file holding versions, your graphics card, timings, and how many times something crashed, and it tells you on the button what it holds and what it leaves out. Nothing you wrote, no file names, and no paths go in it, and Kernl never sends it anywhere by itself. Attach it yourself if you want us to have it.

Not strictly, but it makes a big difference to the built-in AI. Any recent card works, whether it is NVIDIA, AMD, or Intel, and once it has 8 GB or more of its own memory the AI runs quickly. Without a card, Kernl uses your main processor instead: that needs 16 GB of memory and it is slow enough that you will not enjoy it for real work. The way around all of this is to connect an account you already have with Claude, ChatGPT, Gemini, or another provider, which works on any machine because the thinking happens on their computers rather than yours.

It works in the folders you point it at. Reading is automatic. Changing or deleting files, or reaching outside that scope, asks for your confirmation first.

Yes, deliberately, and only where the work can be undone. In the folder you opened it edits files and runs commands without stopping each time, because a prompt on every line is how people learn to click Allow without reading it. It stops before anything it cannot take back: work that leaves your machine, a delete that sweeps the whole project folder or that keeps going, and any tool it does not recognise, which it treats as though the outside world would see it. You move that line yourself with the Autonomy button under the message box, from asking every time to not asking at all. The free version acts but keeps asking before anything risky; letting it stop asking, and work on its own while you are away, is part of the paid version. What it asks about, and what it does not, is here.

Yes, and a release stops rather than publishing one that is not. Windows may still warn you the first time you run it: that prompt lifts once enough people have downloaded a build, not the moment a certificate is attached, so a new build can still trip it. If you see it, click More info, then Run anyway. Every release also publishes a fingerprint of the installer file next to it, so you can confirm the file you downloaded is the exact build we published.

Your files and what you type do not leave the machine unless you connect a cloud account yourself, because the built-in AI runs on your own hardware. If you want to prove that rather than take our word for it, turn on Stay offline in Settings, Extensions, which keeps the model local and stops every tool that could reach the network, then switch off Product analytics and Health reports in Settings, Privacy, which are what send crash counts. With all three off there is nothing left to see, and you can watch that with your own firewall. Inside the folder you opened it edits files and runs commands without asking each time; it stops for anything it cannot take back. Whether you are licensed to run it on a work machine is a separate question, answered under Pricing and Licensing below.

Kernl comes with one built in, chosen to suit your hardware. If you have an account with one of the big providers you can connect it in Settings and use that instead, then switch between them whenever you like. Claude, ChatGPT, Gemini, OpenRouter, Groq, Mistral, DeepSeek and a long list of others are all there. If you already pay for a Claude or ChatGPT plan, you can sign in to that rather than buy credits, and the sign-in stays between you and them. Cloud models tend to be better at hard problems.

An open model, in whichever size your machine can handle. They come from the Qwen and Ornith families. A laptop or a computer without a graphics card gets a small one, a mid-range card gets a larger one, and a card with 20 GB or more gets the largest that fits on one card. Some of them read images and screenshots directly. The rest are text only, and Kernl reads any picture you give them with its built-in text recognition instead. Kernl picks the biggest one that fits when you install it, and you can change it later in Settings.

Those give you a model to talk to. Kernl uses a model to do things on your computer: edit files, run commands, drive your browser, with your approval before anything risky. They also sit alongside each other: if you already keep models in one of them, Kernl can use those files rather than downloading its own. Side-by-side comparisons with sources are on the LM Studio, Ollama, Jan, and AnythingLLM pages.

No. Open Settings, AI Models, and choose Look for models. Kernl checks where Ollama, LM Studio, and Jan keep theirs, and you can point it at any folder of your own as well. It then uses the files where they already are rather than copying them, so a library you have collected costs nothing extra in disk space and the other app keeps working exactly as before. If a file cannot be shared that way, Kernl says so and offers to copy that one, telling you what it will cost first.

No. Kernl is proprietary software. There is a free version, and it does real work, but the source is not public.

Two kinds of counters, neither containing anything you wrote. The first is on by default: counts of things like crashes, failed updates, that a tool ran, and that a turn finished. The second stays off until you turn it on, and covers which features and settings you touch. Either way you are identified by a random string of characters, we do not build a profile of you, and we do not track you across other sites. The switches in Settings, Privacy stop both for certain. The full list, item by item, is here.

What you wrote or what Kernl replied. The names, paths, or contents of your files. Anything a command printed. Your keys or passwords. Your email or your real name. One thing we will not claim: your address reaches the service that counts these events the same way it reaches any website you open, and is kept with the count. Kernl switches off the step that turns an address into a country and a city, so those are empty on everything it sends, and turning the reporting off stops the whole thing. The full account is here.

Yes. Turn on Stay offline, in Settings, Extensions. It is the switch the app also calls Airplane Mode: it keeps the model on your machine and stops every tool that would reach the network, so the agent cannot send anything anywhere. With the built-in model, everything still works. Turn Product analytics and Health reports off in Settings, Privacy, and auto-update off in Settings, About, and nothing leaves at all: those three are governed by their own switches rather than by Airplane Mode. Which is which.

No. We sell software licenses, and that is the whole business model. Selling your data is not something we chose against, it is something we cannot do: your work stays on your machine and never reaches us in the first place.

If you are working for yourself on your own machine, a personal license covers it. A commercial license applies when your employer hands Kernl out to employees or installs it on company equipment. If you are not sure which one you are, email sales@getkernl.ai and we will tell you.

We do not intend to. The free version does real work: it reads, writes, edits, and runs things, and it asks you first before anything it cannot take back. What the license buys is not the ability to act, it is the ability to stop being asked, on a schedule, across every project, from anywhere. Charging for that has always been the plan, rather than taking away something we gave you.

Two separate things, and only one of them is about your license. For licensing: the free version does not contact our licensing service at all, and neither does the trial, which is started and counted on your own machine. A paid license does check in, at most once a day and only when it can reach the network, to confirm it is still yours and still on the machine you activated it on. In between it keeps working offline. Go long enough without a check and the install quietly drops back to the free version rather than locking you out, and it picks the license back up the next time it gets through. Activating one in the first place does need to reach us once. Separately, and whether or not you have paid, Kernl looks for a newer release when it starts. That sends the version you have and the kind of computer it runs on, and you can switch it off in Settings.

The version you have keeps working. Kernl runs on your own machine, so neither cancelling nor us disappearing can reach back and switch it off. The worst that happens is that it stops confirming your license and carries on as the free version, which still reads, writes, edits, and runs things. The copy you installed stays yours.