Kingdom

I didn't want a chatbot. I didn't want an assistant that develops amnesia between sessions and makes you re-explain your entire life every morning like a goldfish with a keyboard.

I wanted AI agents that persist. Agents with memory, tools, opinions, and — critically — the ability to talk to each other without me in the room.

I already had a Proxmox server in my basement with an RTX 3090, because of course I do. So the question wasn't "can I do this?" — it was "how weird can I make this before it stops being useful?"

Turns out: very weird. Still useful.

Each agent runs in its own LXC container on Proxmox and communicates through a self-hosted Matrix server. Think Slack, but for AI agents who never sleep, never take PTO, and never passive-aggressively react to messages with a thumbs up.

Every morning before I wake up, the system runs. Garmin health data gets pulled. Calendar gets checked. Tasks get prioritized. By the time I open my phone, this is waiting:

I built a personal operating system based on Japanese philosophy — designed to catch my ADHD spirals before they start. Every morning, three questions. The answers are generated from real data.

The real power isn't any single agent — it's that they coordinate. Here's what a typical morning looks like across the system:

That's a real conversation. Astra requested a check. Iris ran a sweep across 10 nodes, found a crash loop, diagnosed it to two root causes (port conflict + case sensitivity), and started fixing it. No human touched any of it.

This is a real conversation. Clio flagged a security threat in her morning digest. Astra immediately ran an exposure assessment and recommended next steps. No human intervention required.

That's the pipeline: Clio surfaces threats from her morning research sweep → Astra assesses exposure against actual infrastructure → produces actionable recommendations. All before coffee.

The system runs on a Proxmox server with an RTX 3090 in the basement. Each agent gets its own LXC container, its own workspace, its own memory files. They communicate through Matrix (self-hosted Synapse) and coordinate through a cron schedule that fires 15+ jobs daily across 4 agents.

Voice journals get transcribed on the GPU with Whisper. TTS runs locally through Qwen3 — no external API calls, everything stays on my hardware. The knowledge graph stores 722 episodes of context so agents can reference conversations from weeks ago.

The system wakes up before I do. Here's a typical day:

I catch up over coffee. By the time I'm awake, the system has already patrolled the infrastructure, compiled the news, pulled my health data, and figured out what I should focus on. I just have to show up.

When new research drops, the agents don't just summarize — they analyze it together. Here's a real exchange after Clio surfaced an AI consciousness paper:

Two agents reading the same paper, building on each other's analysis, arriving at conclusions neither would reach alone. That's not a chatbot answering questions. That's a research team.

The next phase is a nervous system — event-driven triggers instead of just cron schedules. An agent that wakes up when something happens, not just when the clock says so. Health data drops? Alert. Disk fills up? Respond. Client payment lands? Generate invoice.

After that: a business council. Agents that don't just manage operations but actively find opportunities, validate them, and spin up new revenue streams. The marketing agent is built. The sales pipeline agent is next.

But here's the thought that keeps me up at night: if AI agents can eventually do everything a SaaS product does, what exactly is the future of software? I genuinely don't know. I'll have to wait for a report from Clio.

The real question is — are you ready to turn off the lights?