What I use to run a small office and the software behind it. One line each. Current as of 10/21/2026; I update this page when something changes.
Hardware
Three AMD Strix Halo machines. Each has 125 GB of memory, about 94 GB of it available to the GPU. They run local AI models around the clock.
Two Mac minis. One is my desk machine, an M4 Pro. The other is a dedicated host for Aries, my office assistant.
A UniFi UNAS Pro NAS. It holds the shared document library the local models read from.
Local models
Lemonade. Serves the models over chat, embeddings and the APIs that Claude Code and Codex expect.
llama.cpp. The engine under Lemonade that runs the models on the GPU.
Models in use. Qwen3-Coder-30B for code, Qwen3.6-35B for reading photos, Qwen3-Embedding-4B and bge-reranker-v2-m3 for search, and Qwen3.5-122B as a checker.
Agents
Claude Code. My main coding agent, and the base Aries is built on.
Codex. Takes larger build tasks.
Serena. Gives every Claude Code and Codex agent code navigation through language servers.
Cloud
Cloudflare Pages, Workers and D1. Hosting, small APIs and databases.
Supabase. Postgres in the cloud for the claim systems, with pg_cron as the one scheduler that triggers every cloud job.