Overnight it goes through leads in your field and checks your website and server. At 7:30 in the morning it sends what happened to your phone. During the day you type or say what you need — and it gets done. You only approve.
Four things that run for me every day. Not from a brochure — from my own operation, where the assistant finds work for me.
Overnight it goes through portals and listings in your field, drops tenders and noise, and in the morning sends only what is worth contacting. It looks for people who wrote themselves that they need something — not a list of every company nearby.
Every morning it verifies the site is up, the certificate is valid, analytics work and nothing internal is exposed. When something breaks, you are the first to know — not the customer who left for a competitor.
Today's calendar, new leads, website status and one piece of industry news with a practical impact. One message on your phone, not five apps. When nothing happened, a short "all good" arrives.
From the car or the building site you record a voice note — in Telegram, WhatsApp or Slack, the interface is up to you. The assistant understands Czech, handles the task and answers by voice. It books a meeting, sets up a Google Meet and prepares the invite, reads your mail and tells you what matters, saves invoices from the inbox into a folder, looks up information, drafts a text.
These are the scenarios running for me. Yours will be different — a real-estate office needs something else than a tradesman or a lawyer. That is what the free half hour is for.
I am not talking about what could be done. This is my own operation — a team of agents, each with its role. One commands, one guards security, one hunts for work, one checks the others. On the left what happens, on the right what lands on my phone.
The agent goes through nine areas — updates, open ports, login attempts, known vulnerabilities. For each one it checks whether it applies to us at all. A false alarm costs trust.
Every client website: availability, certificate validity, analytics, sitemap, and whether anything that should not be public is. A broken website is reported before breakfast.
Two passes: fixed rules with scoring, then an agent over further portals. It drops tenders, job ads and anything off-topic. When it finds nothing, it writes "nothing" — better nothing than filler.
It scans the last 24 hours and picks the single item with a practical impact. When there is none, it stays silent.
Security report, website status, hot leads with today's calendar, one news item. All together, so I can read it over coffee.
I say what I need and the assistant gets it done — it understands Czech and answers by voice. It books the meeting in the calendar, sets up a Google Meet and prepares the invite, looks up information, drafts a text. One voice note, several tasks.
It reads what came in and tells me what deserves attention. Invoices from attachments are saved into a folder by month, a reply to an inquiry is drafted. Nothing is sent — it waits for my "yes".
The morning check found pending updates? The assistant installs them — but only after my approval. It never changes anything on its own.
The whole operation costs a few hundred crowns a month. The assistant does not run around the clock — it starts, does the task and stops.
🟢 Server fine. No failed login attempts, firewall active.
🟡 12 updates pending — none critical. Shall I install them? Reply "yes".
All websites up. Certificates valid, analytics working.
Note: one domain's certificate expires in 9 days — it renews itself, I am only watching.
Law firm, Plzeň — looking for a new website, wrote it themselves, no budget given. Score 64.
Joinery, Beroun — no website, Facebook only, active listings. Score 58.
Today: 10:00 meeting, 15:30 call. I sent nothing — waiting for you.
One thing with a practical impact today — and why it concerns you. When there is nothing like that, this message does not arrive.
"Meeting with Novák tomorrow at half past seven in the evening over Meet, send him the invite and remind me."
Booked: tomorrow 19:30 – 21:00, Work calendar, "Meeting — Novák", Google Meet link in the event. Reminder one hour before.
The invite is ready — I will send it once you reply "yes".
14 e-mails since morning, 3 deserve attention: a website inquiry (reply drafted), a question about a date, a reminder from the accountant.
2 invoices saved to Invoices/2026-09. The rest are newsletters.
It is not one chatbot that does everything halfway. It is a team of specialists with a shared memory — and a scheduler that remembers what worked last time.
One commands and talks to you. Another guards security, another hunts for work, another writes, a counsel watches outreach rules and a reviewer tests the others' work. When one makes a mistake, another catches it — exactly like a team of people.
Everything the team learns sits in one common memory that every agent checks before starting a task. One agent's lesson is everyone's lesson. Your prices, rules and customers are there once — not ten times, each slightly different.
A scheduler with its own memory. It remembers which sources brought a lead last time and which were dead ends, and adjusts what it does next. When you correct it, it writes that down and does not repeat it. That is the difference between automation and an assistant.
This is the difference between an assistant and a toy. Every rule is written into it and cannot be bypassed — not even when it would be faster.
E-mails, offers and server fixes wait for your "yes". The assistant prepares, you decide.
‹ › on the ears, dots on the chest, arrow keys, or drag across the head
The scope varies, so I quote after the consultation — in writing and up front, same as with websites. Nothing gets built until you know what it costs.
Free, by phone or video call. You tell me what slows you down and what you would like to happen on its own. I tell you what makes sense and what does not.
In writing: what the assistant does, when, whom it reports to and what it must not do. With the price. You approve, or we adjust.
The pilot runs for two weeks and during those two weeks we tune it together: you receive the messages on your phone and tell me what fits and what does not.
I train you and the people who will work with the assistant. I offer monitoring and support in operation — and we keep extending the assistant as your needs grow.
For those who want to know how it holds together. This is no no-code toy — it is a team of agents on its own server with clear rules about who may do what.
Every agent is a separate Unix account with its own role in CLAUDE.md. The master commands via sudo -u, the others exchange messages through inboxes. Who may do what is defined by account permissions, not by the model's good will.
Routine runs on cron, learning on a scheduler with memory. Deterministic scripts (website checks, lead scoring) need no AI and cannot hallucinate. An agent steps in where judgement is needed — and it has persistent storage where it remembers what worked last time.
A shared brain: knowledge in linked Markdown, RAG over Postgres with pgvector, connected through MCP. One agent's lesson is available to all. When an agent annoys me by mistake, it gets written down and does not happen again.
The human interface is a chat you already use — Telegram, WhatsApp or Slack, for instance; there are many more options, from e-mail to a custom app. Mine is Telegram: a single authorised account, voice through ElevenLabs in Czech both ways. The server is a small VPS running Ubuntu with ufw, fail2ban, HTTPS via Let's Encrypt and a nightly audit that is read-only — it fixes things only after approval.
My runtime is Claude Code — but it could just as well be ChatGPT (OpenAI), Gemini, Mistral or Llama on your own hardware; the options are countless and the choice depends on the task, the cost and where the data may live. Agents do not run around the clock; each task starts, finishes and exits. You pay for work, not for time.
Tell me what slows your business down or what you would like to happen on its own. I reply within 24 hours and we agree on a time.
