Thomas: tomadigital Cockpit, "Das Kunden-Dashboard einer Ein-Personen-SEO-Agentur“

Right now I am adding new tools and agents and streamlining workflows. In day-to-day practice it looks like this: in the morning the agents have already pulled the numbers from GA4, Search Console, and Google Ads and checked whether anything looks unusual or a service is down. If everything is fine, I get no email and start straight on with client work.

I build myself tools that make my everyday work as a one-person agency easier, and in the simplest way possible. I do not want to sell any of this. Bringing yet another SaaS tool to market is, in my view, very difficult today. A tool of my own that fits my workflows exactly, by contrast, helps me from day one.

Without AI coding, the cockpit would not exist. I describe what annoys me, Claude Code proposes a path and builds it, I test on screen. One example is the watchdog: an agent that checks my clients' traffic every morning and gets in touch when something changes unusually. The code for that was written quickly. What took longer was the question of when a traffic drop is real, because traffic fluctuates normally from weekday to weekday. Now it compares each day with the same weekday of previous weeks and only reports when something clearly diverges on three out of five days. I formulate those rules; Claude implements them.

Next it moves more into the generative side. Image creatives for ads already come from the cockpit today; videos are meant to follow.

My harness is Claude Code in the terminal. I choose the model by task: Haiku or Sonnet for smaller things, Opus most of the time. Many prefer to work in the Claude app, for example with artifacts. I prefer the CLI: one terminal window, nothing that distracts. That fits my approach of keeping everything as simple as possible. Inside the cockpit itself I call Claude Sonnet via the API, for example for work packages and brainstorming, plus Gemini for image motifs.

The stack is deliberately kept simple. Node.js with the standard library, no framework, no build step. The data lives in JSON and JSONL files that I can open in a text editor if needed. Saving goes through a temporary file so a half-written file never appears. For a tool with exactly one user that is enough, and Claude understands it immediately.

I use MCP, skills, and sub-agents — all three. MCP servers, some built myself, connect accounting, Google Ads, Analytics, Search Console, and SEO data. Skills cover recurring workflows, for example SEO audits or crawl reports. I use sub-agents selectively: for a code review with a fresh perspective, or as an advisory board of five personas that challenges me on business decisions. What helps me most, though, is a CLAUDE.md per project and a memory of rules that grew out of mistakes.

For beginners all of this sounds like a lot of jargon. But you can get into it relatively well; in the end they are mostly new vocabulary.

The roughly 20,000 lines contain several months of work. I do not sit on it every day; the cockpit grows with my everyday life: when something bothers me, I change it. That is why it is hard to put into hours. Some evenings and nights have gone into it as well, simply because the topic interests me. For me it is also a hobby and a bit of entertainment.

Billing is not per token but via the Claude Max subscription, and I adjust that to my workload. In phases when a lot is going on and I have the headspace to build, I need Max 20x. In quieter times Max 5x is enough. On top of that come smaller API costs because some workflows use the interfaces directly. Those are around 25 euros a month. For what the cockpit takes off my plate, that is little.

I have been building websites since 2003, when I went freelance. At heart I am a web designer and front-end developer: I have used HTML and CSS continuously since then; I feel at home in websites. On top of that come ten years at SEO tool provider XOVI, six of them as head of marketing. That helps me on every project today: from 23 years of SEO I know which numbers really matter. At XOVI I learned how a tool is conceived and what users need from it. And as a front-end developer I read what Claude builds, see immediately when a page renders wrong, and can say exactly how it should look. I am a generalist and self-taught, and for people like me this is the perfect time: anyone who understands enough across many areas can, with AI, build things today that used to require a whole team.

I have been working intensively with Claude and AI coding since early 2026. Since April the cockpit has been emerging that way, and by now a large part of my agency runs on self-built tools and agents.

That background is not a prerequisite, though. Many never approach their own tools because they believe they lack the technical skills. My advice: start small. Instead of planning a big MVP right away, better solve one thing that bothers you in everyday work and see whether it really helps. The cockpit began as a simple client list. Everything else was added only once the value was there. Anyone with a task that annoys them every day has already found their first project.

Hardware is not my field, so I look at it as a user. What I find exciting is that agents are currently moving out of the chat window. OpenAI has introduced Dots agents that run continuously in the background and complete tasks on their own. Meta shows with Muse Charm a small keychain device through which you talk directly to your agent. I see a direction in that: the agent becomes a constant companion; the device is only the remote control.

That agents will soon develop such devices on their own is something I do not yet see. The limit lies in feedback. With my cockpit I see within seconds whether something works, and I can undo it if needed. A device needs prototypes, manufacturing, and tests, and every mistake costs time and material. On top of that: the more an agent handles on its own, the more important the question becomes of who controls it. For me it remains: Claude builds, I decide.

In the end, though, people are still needed — who knows for how much longer. And those people also have to want to operate all of it. Right now everyone with little money has the same chances: some build SaaS tools, others like me build their internal processes and workflows. Still, only a small share uses these opportunities. SEOs have always been the nerds of the internet, a small group that enjoys this kind of thing. With AI it is similar: only a fraction of users really build something themselves. And even when I explain it to others, only a fraction of them start themselves. The interesting question for the future for me is therefore: How does AI manage to bind people to itself?