Field Notes · No. 017 min read

The Real Work Isn't in the Chat Box

Most people meet AI as a text box that answers questions. The bigger payoff comes from connecting it to the systems you already run. Here is how I connected two monitoring platforms and built a live dashboard of my home lab without writing a line of code by hand.

Ask most people what they do with AI and they describe a chat box. They paste in an email and ask for a better version. They ask a question and get an answer. That's useful, and it's also the smallest slice of what these tools can do.

The work that changes how a business runs looks different. The AI isn't answering questions. It's reading data out of one system, making sense of it, and putting something useful into another. Most organizations don't lack software. They lack connections between the software they already pay for, and someone usually fills that gap by hand.

I see this every week in client work, but I can't walk you through a client's systems. So for the first issue of Field Notes, I'm using my own: the house, the servers I run at home (my "home lab"), and a dashboard I've wanted for years. The shape of the problem is one most businesses will recognize: two systems that don't talk to each other, a spreadsheet someone maintains by hand, and no developer on staff. I didn't write any of it by hand.

First, what is Home Assistant?

Home Assistant is open-source software for running a smart home. It runs on a small computer in my house, not in a vendor's cloud, and connects lights, locks, thermostats, the garage door, and hundreds of other devices and services. Each one shows up as a live reading or control: a light that's on, a room that's 72 degrees, a door that's locked.

With everything in one place, you can build on top of it: automations ("turn off every light when the last person leaves") and dashboards that show the whole house at a glance. Mine tracks about 2,200 readings and controls. When a device or service isn't supported, someone can write an integration, a small add-on that teaches Home Assistant to talk to it. The two I'm writing about today didn't exist until a few weeks ago.

The setup: Claude Code, MCP, and house rules

The tool doing the work is Claude Code, Anthropic's AI agent that runs in a terminal. Unlike a chat window, it can read and write files, run commands, and use the tools I give it access to.

One common way to give it tools is MCP, the Model Context Protocol, an open standard for connecting AI agents to other software. Home Assistant has an MCP add-on that runs inside it. Once Claude Code is connected to it, Claude can do what I'd normally do by clicking through screens: look up devices, check history, install integrations, build dashboards, restart the system.

  Home lab spreadsheet ──┐
                          │
  Zabbix API ─────────────┼──▶  Claude Code  ──MCP──▶  Home Assistant
                          │          │
  Project notes ──────────┘          └──▶  New integrations (open source)
The setup. Claude reads from everything on the left and builds into everything on the right.

The connection is half of it. The other half is instructions. I keep a project folder for the house with a plain-language file that tells Claude how this house works and what the rules are. The most important rule: never edit Home Assistant's configuration files directly. Always go through the tools. That keeps every change going through the same front door I'd use, where I can see it and check it.

Enter Zabbix, a very different system

The house is only half my setup. The other half is the home lab: a rack of servers, a firewall, network storage, a couple of GPUs for running AI models locally, and about 30 self-hosted web apps. To keep an eye on all of it, I use Zabbix, a monitoring platform built to watch servers in data centers. Mine watches 48 machines and more than 8,000 individual readings.

Zabbix and Home Assistant organize information differently, name things differently, and expose it through very different interfaces. Getting them to understand each other used to mean weeks of reading technical documentation.

Zabbix doesn't have an MCP connection in my setup, and it didn't need one. Claude read Zabbix's API documentation, wrote its own scripts to talk to my server, and got to work. MCP makes connections easier, but the real requirement is simpler: the system has an API, and you're willing to give the AI controlled access to it.

The bridge: ha-zabbix

Home Assistant ships with a basic Zabbix integration, but it couldn't do what I needed. So I described the outcome to Claude. Every Zabbix machine should appear in Home Assistant with its status and open problems. The Zabbix server's own health should be visible. I should be able to choose exactly which readings come across. And when something breaks, Home Assistant should know, so it can flash a light or send a notification.

Claude wrote it. The result is ha-zabbix, an open-source integration anyone running Home Assistant can install. It went from first commit to version 0.3.0 in nine days: about 4,800 lines of Python, none of which I typed, plus about 4,000 lines of automated tests. Those tests run against Zabbix's official server software, so when Zabbix ships an update, I can find out quickly whether anything broke.

On my system, ha-zabbix created 49 devices and 308 readings. Claude also installed it, configured it, and ran the test plan through the Home Assistant MCP tools, writing up the results as it went.

A companion, ha-zabbix-alerts, lets Zabbix push problems to Home Assistant the moment they happen instead of waiting for the next check. In testing, a new problem arrived in 0.8 seconds. It's installed, but I haven't switched it on for everyday use yet.

I didn't write a line of Python. I didn't write a line of YAML. I never learned the Zabbix API. I wrote down what I wanted and checked that I got it.

From a spreadsheet to a dashboard

With the bridge in place, the real project could start. I had a spreadsheet of my home lab: every machine, its address, what it does, what operating system it runs, and which network it sits on. Thirty rows. Most teams keep a list like this somewhere.

I gave Claude the spreadsheet and asked for one dashboard showing the health of everything on it. Here is what it did:

  1. It found the gaps. Claude compared the spreadsheet to what Zabbix was already monitoring and what Home Assistant could already see, then wrote a short spec of the missing data the dashboard needed.
  2. It added the monitoring. It wrote a script that created the missing checks in Zabbix, covering server load, storage space, network health, and whether each of about 30 web apps was up. Those readings flow straight into Home Assistant.
  3. It built the dashboard. Eight sections, from infrastructure and compute to media, backups, and remote access, with 41 service cards. Each card takes its color from the problem severity Zabbix reports, so a glance tells me what's fine and what isn't. The built-in cards couldn't do that, so Claude wrote a custom one. It set up the whole dashboard through Home Assistant's own tools, not hand-edited files.

That's three sources on one screen: a spreadsheet, a monitoring platform, and a home automation platform. I supplied the inventory, decided which readings mattered, and reviewed each version until it was right.

A tool I needed: ha-ai-usage

Once connecting systems gets this easy, you start noticing gaps you used to live with. Mine was AI subscriptions. My Claude, ChatGPT, and Google AI plans each have usage limits that reset on their own schedules, and running out halfway through a client deliverable is a bad afternoon.

In two days, Claude built ha-ai-usage. It checks my current usage with each vendor and shows it in Home Assistant: how much of each limit I've used, when it resets, and whether I'm on pace to run out first. The projection learns my usual weekly rhythm from history. It knows Tuesday mornings are heavy and Saturday nights aren't, and it reports how confident it is.

Both ha-zabbix and ha-ai-usage are open source on GitHub if you run Home Assistant and want to try them.

What I actually did

It would be easy to read this as "AI did everything." That's not quite right, and the difference matters.

  • I knew what I wanted. Every one of these projects started with a clear description of the outcome, written in plain English.
  • I supplied the context. The spreadsheet, the house rules, the list of what matters. The AI was only as good as what it knew about my environment.
  • I made the judgment calls. Which data matters, what "broken" means, which shortcuts are acceptable and which aren't.
  • I checked the work. I looked at every dashboard, read the test results, and sent things back when they were wrong.

None of that requires knowing Python or the Zabbix API. All of it requires knowing your own systems and what good looks like. Most business owners and managers already have that.

Why this matters outside the home lab

Swap out the nouns. Instead of Home Assistant and Zabbix, it's your accounting system and your dispatch software. Instead of a home lab spreadsheet, it's the one someone on your team updates by hand every Monday. Instead of a dashboard on a wall tablet, it's the report your managers wish they had.

The pattern is the same: systems you already pay for, data that already exists, and a missing connection between them. Building that connection used to mean hiring a developer or waiting on a vendor. In my own work, it now takes a clear description of the outcome, the right access, and someone who knows what good looks like.

There are real cautions. An AI with access to your systems needs the guardrails you'd give a new employee: limited permissions, clear rules, and changes someone reviews. And someone still has to own the result after it ships.

That's the work I do with teams. Not selling another AI product, but helping them connect the tools they already have, then build and maintain those connections themselves. The code isn't the hard part anymore. Knowing what to ask for is.

Who Wrote This

Nick Card

Practitioner, not evangelist.

I'm Nick Card. I studied engineering at Harvey Mudd, spent years as a patent engineer in Silicon Valley, and now run operations for my family's trucking company: the shop, the yard, and 500 trucks and drivers. That job keeps me honest about what AI can actually do in a real business.

Through NKCX, I help business owners and leaders put AI to practical use with consulting, hands-on coaching, and the AI for Business Series.

Working on Something Similar?

If you have systems that should be talking to each other and aren't, I'd like to hear about it.

Email Nick