A home AI assistant doesn’t need to be a data center or a sci-fi face on a screen. A sensible server, a bit of patience, and a healthy dose of skepticism toward autonomous software are enough.
Clicked because of the pizza-ordering AI? The assistant can do a lot more 😉
When people hear “home AI assistant,” most picture either a smart speaker that answers the weather with an ad or a sci-fi household where the fridge haggles with the insurance company. Czech maker KutilDomácí took a third route on X: he got OpenClaw running on his own hardware, named it Soptík, and started handing it real tasks.
A good home assistant’s job is to notice problems early—unreachable server, printer offline, failed backup, suspiciously empty fridge.
Not a chatbot—an assistant with tools #
The key difference from a plain chat window is that OpenClaw can use tools. Not just a language model, but calendar, email, smart home, monitoring, files, or scripts on the home server. The model isn’t a source of truth; it’s a dispatcher that knows when to check the real source.
This separation matters. The assistant shouldn’t “guess” whether Home Assistant is running—it should look at the actual state. And it definitely shouldn’t decide on its own to tweak the firewall, order a truckload of pizza, or trigger any other minor digital apocalypse.
Soptík therefore follows a few simple rules:
- Read and analyze freely, but only inside a pre-defined scope
- Infrastructure changes, deletions, and external actions require confirmation
- Keys and passwords live in separate
.envfiles with proper permissions, not in shell scripts - Anything that can change gets verified at the source instead of being confidently invented
It’s less Hollywood. It also greatly reduces the chance that your smart assistant will one day optimize the household into “no internet, but a beautifully clean git repo.”
What it runs on #
No special hardware. According to KutilDomácí, Soptík is a VM on Proxmox running on a regular eight-core Intel box with 32 GB RAM. It gets four cores and 8 GB of its own.
The model itself doesn’t run locally. OpenClaw is connected to OpenAI, so the home machine only handles tools, access, and context. The thinking happens in the cloud.
For makers that means two things: you don’t need a GPU or anything expensive—you can run this on whatever spare hardware you have. Every query still leaves the house, though, including whatever Soptík happens to see on your network.
Home network #
The setup is a mix of routers, switches, NAS, Proxmox host, VMs, containers, Home Assistant, cameras, other IoT gadgets, 3D printers, and a few devices that exist mainly because someone once plugged them in “just to test.” The network was the obvious first project.
Soptík started with an inventory via read-only access to the MikroTik and safe network discovery. It then deployed Uptime Kuma for service checks and NetAlertX for LAN device overview. The result isn’t a sci-fi control room—just a sensible list of what’s running where, what should be reachable, and what can wait after a longer outage.
One dashboard now shows the state of the router, NAS, Proxmox, Home Assistant, Grafana, Zigbee2MQTT, UniFi Controller, Frigate, cameras, and the rest. Soptík was also told not to panic-report every Wi-Fi bulb blip in the first week. A surprisingly rare trait—not just in AI.
A portal instead of a bookmark list #
When you have more admin interfaces than houseplants, you start missing a proper index. Soptík built a simple LAN portal: links to monitoring, dashboards, and service admin pages plus a short inventory. It runs in a small nginx container and is exposed only on the home network.
It’s not a polished app after thirty sprints. It’s a plain static page that solves a real problem: where exactly is Grafana, Frigate, Uptime Kuma, or the Zigbee2MQTT admin page? In home infrastructure, the simplest solution often gives the best ratio of usefulness to future swearing.
The Telegram prompt was one sentence long.
And the pizza? #
Peak home automation can’t be just monitoring. One day pizza came up. Soptík was told to browse the menu, build an order from the given preferences, and send it off.
Pizza selection is deceptively sensitive: allergies, kids’ portions, half-and-half, unexpected pineapple escalation. And payment. Soptík never got those rights, so the order ended up as cash-on-delivery.
It worked. The family picked up the pizza, the bill wasn’t painful, special requests went through, and Soptík never touched the household budget. Decent first day on the job.
My take: #
OpenClaw isn’t interesting because it could replace every app. It’s interesting because it can connect them and keep context across them: it knows what’s on the home network, what should be working, what the priorities are, and when it should ask first.
The best home AI assistant isn’t the one with the flashiest tricks. It’s the one that’s useful when something breaks, doesn’t create more damage than it fixes, and occasionally helps solve the truly critical tech problem: what’s for dinner. 🧯
Chiptron adds #
Basic security measures
Soptík’s rules point in the right direction and are worth tightening. If you want to build something similar, here’s a checklist I recommend running through before you give the agent access to its first device.
- Put it on its own VLAN and control where it can reach. The agent runs in its own network segment and the firewall only allows what it actually needs: the model API, package repos, and the specific device addresses it monitors. Nowhere else. If something goes wrong, it has nowhere to spread and no one to send anything to.
- Give it its own accounts. Its own user on the router in a read-only group. Its own token in Home Assistant under a limited user. Its own SSH key on the server. Never hand it the credentials you use yourself. If something breaks, you revoke its access, not yours.
- Set rights in the service, not in the prompt. Telling the agent “don’t change anything” in instructions is like taping a note to a cabinet that says “do not open.” The model also reads text none of us wrote: device names on the network, web pages, emails. And text can convince it that this time is an exception. If it should only read, that has to be enforced on the router itself.
- Don’t let it see passwords at all. Instead of a token in a config file the agent can read, write a small script that keeps the credential inside itself and only let the agent run that script. It says “check router status,” the script logs in on its behalf and returns the result. The password is never visible to the agent, so it can’t accidentally paste it into chat.
- Give it buttons instead of a shell. When an agent can run any command, it can do anything it thinks of. If you instead give it a handful of ready-made functions like “restart service” plus a list of services it may choose from, it simply can’t do anything else. You lose a bit of convenience and gain certainty.
- Irreversible actions must be confirmed. Deletion, restarts, anything sent outward, spending money. And make the confirmation show exactly what it’s about to do. Not “really?”, but “delete snapshot from September 3rd.”
- Keep cameras, backups, and money separate. It has no business with payments. Backups should be read-only at most— an agent with delete rights is a fast way to lose everything at once. For cameras, a reasonable compromise is to give it access to events (Frigate saw motion at the gate at 14:12) but not the actual video feed. Your family’s video has no reason to leave for someone else’s infrastructure.
- Log somewhere the agent can’t write. Record every tool use on a different machine or at least in a file the agent can’t touch. When it eventually does something unexpected, you need to trace back what it read just before. The answer is usually there.
- First two weeks: only watch. No actions, just answers questions about the network. You’ll see what it hallucinates and where it gets things wrong before you let it change anything. You’ll also tune what it should actually report. An assistant that beeps at every trivial thing stops being read after a week.
None of this is complicated. Most of it is an afternoon’s work and saves you a very boring evening sometime later.
Alternative: Hermes Agent
In the thread discussion another name keeps coming up. Several comments mention Hermes Agent from Nous Research, so here’s a quick look.
Hermes Agent is an open-source self-hosted assistant under the MIT license with the same core idea as Soptík: model plus connected tools. It supports persistent memory that learns your projects, task planning from plain-language requests, delegating work to sub-agents, web browsing, vision, image generation, and text-to-speech. It works from Telegram, Discord, Slack, WhatsApp, Signal, email, or the command line. You can deploy it with a single curl command or as a native app on macOS, Windows, and Linux. Execution is sandboxed and five backends are available: local, Docker, SSH, Singularity, and Modal. You choose the model; over three hundred are accessible through the Nous Portal.
Links: KutilDomácí on X | Hermes Agent | Hermes Agent on GitHub