Run a homelab from one dashboard with a local LLM agent

The poster, a software engineer, wanted a single place to manage the many self-hosted services in their homelab and keep it private, so they built Valet and run it on their own hardware. It is a dashboard with a chat agent that can use a local or hosted model; at home it runs Qwen 3.8 27B on an RTX 5090. The agent has tools for 185+ integrations and custom plugins, so it can be asked to pause Pi-hole, look into crash logs or request a movie. It also reads its own docs, so asking for a change to the dashboard makes it edit the config file, which waits for the owner's approval. Docker access is read only through a socket proxy. The poster says Claude wrote most of the code while they designed and reviewed it.

On the 5090 the model generates about 230 tokens a second and most replies take 2 to 5 seconds. Qwen passed roughly 90% of a small eval set the poster wrote. It does worst on open-ended troubleshooting such as working out why a timer is not running, where it can burn through many tool calls. Tool definitions were once about 90% of the prompt; hiding unused tools and loading tools lazily cut tokens per answer by 30 to 50%.

Turning the model's thinking off made it 2 to 3 times faster but it began inventing answers to follow-up questions. With many services the tool list alone can crowd the context, so trimming tools matters more than the model choice. The dashboard is read only until auth is turned on, API keys stay on the server and are never shown to the model, and any change waits for approval. The poster calls it early and expects rough edges.