Translated from the German original. Read the original in German.
Since February I have been building my personal AI assistant with OpenClaw – and I am more enthusiastic about it every day.
From the outset my goal was to set up a professional personal AI assistant that takes a lot off my plate as a consultant – email management, scheduling appointments, document management with direct access to Nextcloud, scanning and assessing public tenders. The learning curve was steep – but that is exactly what made it fun.
One of my most important insights was that structure beats intelligence! Clear skill rules that tell the agent how to think, hierarchical memory files that preserve context across sessions, cookbooks for Python, shared libs and skills for recurring workflows – and heavy use of Python for cron jobs that keep everything running.
At the heart of it is OpenClaw on a Mac mini M4 Pro with 64 GB of unified memory. Several language models run locally via Ollama at the same time – from small, lean models for simple tasks to 30B models for complex classification. Another insight was that, with limited RAM, you need a dedicated coordination mechanism for switching models. I built one myself as a lease hub – since then everything has run smoothly, because the workflows automatically switch to the model best suited to each workflow.
A selection of my workflows:
🧾 Incoming invoices – incoming emails are processed automatically overnight, invoices are classified by AI and filed as PDFs in Nextcloud, and I receive a structured summary via Telegram.
📰 Morning news – every morning at 6 a.m. an RSS collector aggregates articles from ten sources, a local LLM writes short summaries and the digest lands straight in my Telegram channel.
📅 Appointment management – calendar invitations sent by email are detected automatically and imported into my Nextcloud calendar, and I get timely reminders with travel time and context. All via CalDAV, fully local.
⚡ Teleseo meter readings – every month a Playwright script automatically reads the current electricity meter data from the FRITZ!Box and submits it straight to the portal. It just works.
🏛️ Tender radar (UVgO & TED) – every morning my system automatically searches dozens of public procurement platforms in Germany as well as the EU-wide TED database for relevant tenders. What used to be hours of manual research now runs fully automatically.
My verdict after three months:
A fully sovereign platform. My data stays local, my processes stay under my control. And when I need the power of the big American language models, I connect them via API – but I don’t have to.
Digital sovereignty doesn’t have to be a major project. You can start small. And it pays off.
