Home AI server

Problem

My AI costs were all API calls - every image and every agent reply went out to a provider and came back on a bill. I wanted to move the parts that didn't need a frontier model onto my own hardware, and learn how local hosting actually works while doing it. Two earlier attempts on a gifted desktop and an old laptop went nowhere, because neither had a graphics card worth using.

What I built

A headless Ubuntu server built from a second-hand gaming PC, hosting my photo library, local image generation and three AI agents I talk to through Telegram. Each agent has one job and only the tools it needs — one runs the server, one handles image generation, one puts together a morning digest. The one I lean on most watches the server itself: a shell script gathers the facts every fifteen minutes and only wakes an AI model when something is actually wrong, so I get a message when something breaks and silence the rest of the time. I specced it out and wrote the code with Claude Code. Early on the admin agent tried to give itself root access and was blocked, so now anything needing root goes through me at a terminal instead of through an agent.

Tech used

• Ubuntu Server 26.04
• OpenClaw
• xAI Grok 4.5
• Ollama
• Qwen3 14B
• NVIDIA CUDA
• Tailscale
• systemd
• Cockpit
• Bash
• Python
• Telegram Bot API

Result

Image generation now runs on my own graphics card, so that part of the bill is gone — though the agents still use a paid model for the thinking. It's been running since July, and the clearest catch so far was memory use hitting 93% because Forge was holding far more RAM than it should have. A restart cleared it and I never found the root cause, but I wouldn't have noticed until the machine started struggling.