Running an AI agent locally can keep financial records, embargoed product data and other sensitive files off cloud services, but a hands-on test shows that privacy does not remove setup work or reliability problems. The Verge installed Hermes Agent on a Mac Studio with 256GB of unified memory and selected a 125-billion-parameter Qwen model occupying about 105GB.
The system delivered a daily briefing through Telegram, reorganized a Steam library and analyzed private records. Sorting more than 400 games required a Steam web API key, which the tester revoked after the one-time task. A more ambitious effort to automate repeated laptop benchmarks remained unfinished.
Local operation avoided per-token charges and made private-data tasks acceptable to the tester. It also required expensive hardware: the Mac configuration cost about $12,000, while a four-machine demonstration cluster approached $50,000. Smaller models and devices are planned for later testing, so those prices are not minimum requirements for every local workflow.
The trial also exposed ordinary operational limits. A scheduled briefing failed when macOS was asleep, and even powerful local models did not make automation effortless. Local AI changes where data and computation live; users still have to manage credentials, permissions, uptime, model choice and verification of the agent’s work.