This is the part that did not exist five years ago.
Setting up a filtering gateway used to mean knowing nftables, DHCP, DNS and container networking. Real skills, and not ones you pick up in an evening. That is exactly why the only realistic option for most families was a subscription that watched their children on somebody else's servers. The complexity was the moat, and the moat is what you were paying rent on.
An AI coding agent closes that gap. It reads the repo, works out what your particular hardware needs, writes the configuration, brings it up, and tells you in plain English when it wants you to plug a cable in. You stay in charge, you approve what it does, and you can ask it what any of it means at any point. It is a patient expert who does not mind being asked the same question twice.
So the complexity did not go away. It stopped being yours to type. What you are standing up is genuinely the sort of thing a company pays a network engineer for: an isolated segment, a firewall that rebuilds itself from a database every fifteen seconds, DNS with a different policy per person, containers, a database, timers, and a test suite that fires real packets to prove the safety properties hold. You should want that to be complicated. A parental control system that was simple enough to install in two minutes would be simple enough to get wrong in two minutes.
The only way to have a world-class appliance like this in your house for no money is to do a little bit of work. The good news is that the work you do is reading and deciding, and the agent does the rest.
Any of these will do the job.
They all read a repository, run commands with your approval, and talk in plain language. If you already pay for one of these, you are already set. Check current pricing on their own sites, because it moves.
Claude Code what Genkan was built with
Anthropic's terminal agent. Genkan's own guides are written for it and tested with it, so it is the smoothest path. Included with a Claude Pro or Max subscription, or pay per use with an API key.
npm install -g @anthropic-ai/claude-code
OpenAI Codex
OpenAI's coding agent, included with a ChatGPT Plus or Pro subscription. If ChatGPT is the one you already pay for, start here rather than adding another bill.
npm install -g @openai/codex
Gemini CLI
Google's agent, and the one with a usable free tier: sign in with an ordinary Google account and you get a daily allowance that is generally enough for a setup session. The cheapest way to find out whether this project is for you.
npm install -g @google/gemini-cli
Already paying for one? Use that one. There is nothing to gain from a second subscription. Genkan's guides are written in plain Markdown and describe what needs to happen, not which agent has to do it. Any capable agent can follow them.
Four moves. Twelve prompts. A few evenings.
We used to say "four steps and you only type the fourth". That was true in shape and a bit glib in practice, so here is the honest version.
Twelve rather than one, because the steps that deserve to be separate really do deserve it. Working out which network port is the kids' side is its own prompt, because getting it wrong takes the whole house offline. The database is its own prompt, because it is the one part no script in the repo does for you. The access point is its own prompt, because a router that is still bridged to your house LAN is the single most common reason a build stalls, and the gateway will refuse to start rather than fight your router.
Each prompt is self-contained, so pasting it into a fresh session still works. Each names the files the agent must read first. Each tells you what you should see and what to do if it looks wrong. And every one of them ends by telling the agent to stop and ask, rather than to guess.
That last line is not decoration. An agent trying to be helpful will sometimes "fix" a failing safety test by relaxing the thing it tests, and on this particular box that thing is the wall around your children's network. The prompts are written to make that behaviour impossible to mistake for success, and there is a final prompt in the file that audits for exactly it a few days later, ideally with a different agent.
What only you can answer
Which network adapter goes to the kids. What the wifi should be called. Who lives here, how old they are, how much time each of them gets, and whose tablet that is. None of it is technical, and all of it is the actual parenting decision underneath the software.
What you cannot hand over
Physical work, and root. Someone has to plug the USB adapter in and cable the access point. And deploy.sh and the test suites need sudo, so if your box asks for a password the agent will stop and hand it back to you. That is correct behaviour, not a failure.
What the agent is not
Part of Genkan. It builds the thing and then it is done. Close it and never open it again and the gateway keeps running, because nothing in the running system depends on it. Most families keep one around because talking to the house in plain English is genuinely nicer than a CLI, but that is a choice, not a dependency.
What people ask before they start.
Why is there no installer? Because writing one that is actually reliable across every distro, every network stack and every USB adapter is a bigger job than the gateway itself, and a half-reliable installer on the box your family's internet runs through is worse than none. So the effort went into the thing that works instead: a repo written to be read by a machine, plain-language runbooks, one CLI, and twelve prompts written and checked against what the code actually does. If that changes, this page will say so.
Do I need to understand what it is doing? No, but you should be able to read it back. Each prompt tells you what a working result looks like, so you are checking an outcome rather than auditing a command. The one place to slow down and actually read is the network interfaces: nominate your uplink as the kids' side and the house goes offline, and no software can tell your intent from a MAC address.
Is it safe to let an AI run commands on my machine? Every agent above asks permission before it changes anything, and you can read each command before you approve it. Use the spare box, not your work laptop, and the worst case is that you reinstall an operating system on a computer that was in a cupboard.
Does the agent see my family's data? During setup it sees the repo and your machine's configuration. It is not part of Genkan once it is running. Genkan itself never phones home, and the agent is not sitting inside it watching your household. When setup is done you can close it and never open it again.
What if it gets stuck? Tell it what you see. That is genuinely the whole technique. Paste the error, describe what happened, and let it work the problem. The prompts file ends with a troubleshooting prompt written for exactly this, and it includes the line that matters: if the fix involves weakening the segment isolation, the DNS forcing, the segment guard or the safety net, stop, because those four are never the fix.
What does it cost? A setup session is well within a free Gemini allowance or an existing Claude or ChatGPT subscription. Genkan itself is free and always will be: no account, no tier, no upsell, and nothing to enter a card into at any point. That is not a launch offer, it is written down as a commitment in the privacy charter, alongside no telemetry and no monetisation that depends on data.
I have never used a terminal. Then this is a good first reason to. You will be typing a handful of commands that are written out for you, and asking an agent for help with the rest. Plenty of people meet Linux this way and find the fear was the hardest part.