KeplerClaw

What it is

KeplerClaw is my self-managed AI infrastructure. It runs on an Ubuntu server (Intel i7-8700) in my room, is reachable through Tailscale, and hosts the personal agents, automations, memory, and editorial tools I use regularly.

Why I built it

I was tired of AI tools feeling like scattered subscriptions. I wanted something I could shape, run, inspect, and trust. Infrastructure, not apps. Something where I own the inputs and outputs.

What it does

• A Telegram-based AI assistant I talk to from anywhere, with cross-session persistent memory so it remembers context between conversations.
• Google Calendar integration so it can answer “what does today look like” without me opening a tab.
• Scheduled autonomous briefings that arrive on their own, summarising what matters to me each morning.
• GhostWriter and the news intelligence pipeline run on the same box, so editorial tooling and personal infrastructure share one environment.
• n8n automation workflows for gluing things together without writing a service every time.
• Remote SSH management from any of my devices through Tailscale.

Everything is Docker Compose based, so bringing the whole stack up is one command.

The hard part

The hard part is not any individual component. It is keeping the whole thing stable, observable, and cheap to run. A personal server has no ops team. If something breaks at 3 a.m., it stays broken until I wake up. Designing for that constraint is its own discipline.

What I learned

• Persistent memory is what turns a chatbot into something that feels like infrastructure. Without it, every conversation starts from zero.
• Tailscale is the single best decision in the stack. It removes an entire class of networking pain without introducing a VPN configuration project.
• Docker Compose is enough. You do not need Kubernetes to run a personal server well.

Current status

Internal. Actively used, actively evolving. Not a public product.