Built a $0/month local AI agent on a 32GB mini PC — full setup and pitfalls
Someone with 16 years in facilities maintenance who self-taught AI built a on a mini PC to escape rising SaaS subscription costs and avoid handing data to . The hardware is a GMK K12 mini PC with a Ryzen CPU, 32GB of RAM, and only (no discrete GPU), dedicated entirely to running the agent. The model runs through Ollama, using the gpt-oss:20b model.
The is Hermes Agent, an open-source (MIT licensed) project from Nous Research. The agent reads and writes to an Obsidian vault acting as a 'second brain,' and a dedicated 2TB NVMe drive keeps model files off the main system drive. The core design is a tiered setup: the local gpt-oss:20b model acts as the 'worker,' handling high-volume tasks like filing, drafting, and organizing at zero cost, while an existing acts as the 'manager,' handling harder reasoning and coordinating everything through a one-page state log.
The whole system runs with no monthly fee, on hardware sitting on a desk, and the author notes several undocumented pitfalls that cost days to work through.
Key points
- Built a on a mini PC (Ryzen CPU, 32GB RAM, only) running with zero monthly cost
- Uses Ollama as the model server, running the gpt-oss:20b model
- is Hermes Agent, an open-source MIT-licensed project from Nous Research
- Tiered design: local gpt-oss:20b handles routine tasks, an existing handles complex reasoning and coordination
- Added a dedicated 2TB NVMe drive so model files stay off the system drive