Building my personal AI Assistant

I built an AI assistant that runs on my homelab. I named her Tatl, after the fairy from Majora’s Mask.

In May I attended AI Engineer Singapore, Singapore Foreign minister Vivian Balakrishnan gave a keynote about his personal AI setup, a self-hosted assistant running on a Raspberry Pi. It was his second brain. He shared the setup in a public gist.

It clicked so I had to build my own.

What Tatl Can Do

Tatl connects to the services I use daily:

• Gmail - summarises yesterday’s emails every morning
• Google Calendar - briefs me on today’s agenda each morning
• Garmin + Strava + Oura - pull my fitness data and training plans on demand
• Home Assistant - check cameras
• Homelab - monitors services like Uptime Kuma and Grafana, sends alerts when something is down
• Voice notes - send a voice message and she transcribes it locally via whisper.cpp

The morning briefing means I can mute most of my notifications during the day.

Every message is processed with persistent memory. She builds up a knowledge graph over time and recalls relevant context before responding.

Visualising the Wiki

The memory graph synthesises into markdown pages that Tatl writes and maintains. To browse them I use Quartz - a static site generator that turns markdown into an Obsidian-style website with graph view. Self-hosted privately.

The Stack

Built on NanoClaw.

• Runs on the Mac Mini and handles message routing
• Each conversation spins up an isolated Docker container running Claude
• Credentials are managed through a local vault - containers never see API keys
• Memory is stored in a local knowledge graph via mnemon
• Voice notes transcribed locally via whisper.cpp
• Wiki is visualised using Quartz

What I learned

I tried local models first - Qwen, Mistral, Gemma. At the time, they weren’t reliable enough for tool use. The ecosystem is moving fast though and I will revisit this.

The memory layer is what makes it feel different. Without it, it’s just a chatbot with integrations.

Get your own

NanoClaw is open source and runs on anything from a Raspberry Pi to a Mac Mini.

• Get a Claude API key
• Start with Telegram as the main messaging channel.

Then explore from there, you’ll find yourself adding more of your day-to-day to it naturally.

The hardest part was finding a name for the agent.