Local AI offers a useful trade: keep sensitive work on your own computer, then take responsibility for choosing models, granting access, and getting the setup right. An early test of Hermes Agent suggests the privacy can be worthwhile, though a private assistant is still very much a computer program that needs supervision.
Privacy is the point
The experiment ran Hermes Agent, a free, self-hosted app for macOS, Windows, and Linux, on an M5 Ultra Mac Studio with 256GB of unified memory. Its model picker made it straightforward to install Qwen 3.8 Flash Next, a model with 125 billion parameters and a file size of about 105GB. Running local models also meant there were no token charges for each request.
That much hardware is not a requirement for every useful task. It did, however, let the tester start with a large model while exploring smaller ones for other machines, including an M6 Mac Mini, an M5 MacBook Air, an Asus TUF Gaming A14 with an AMD Strix Halo processor, and RTX Spark systems.
The appeal was practical, not mystical. Financial records and laptop specifications under embargo were information the tester did not want to send to a cloud service. Keeping those tasks on the computer made using an AI tool acceptable where uploading the material would not have been.
Small jobs, mixed results
The first project was a morning briefing. A scheduled task checked email and calendar items, then added a short weather report. It initially failed because the Mac was asleep when the job was due to run at 7:30am. Once that problem was fixed, the briefing ran, although it later broke more than once and still needed better inputs to become genuinely useful.
A more successful job sorted a Steam library of more than 400 games. Hermes could inspect much of the library through the installed Steam client, then organize games by genre while leaving existing personal categories in place. To make changes, it needed a Steam web API key. The task took a few minutes, and the key was revoked afterward.
Hermes also analyzed financial records and prepared a spreadsheet comparing laptop specifications. These were modest jobs, but they demonstrated the particular value of local processing: the information could be used without being uploaded to a cloud service.
Permissions are part of the bargain
An agent that can act on a computer needs access to the systems it is meant to control. In this test, organizing Steam required an API key. Other tasks involved email and calendar data. Local processing changes where information goes, but it does not make permissions irrelevant. Users still need to decide what an agent can access, and remove access when it is no longer needed.
The tester is also working on a larger project: a script that could automate laptop benchmark runs. The existing process involves running a dozen or so tests three times to calculate an average. Teaching Hermes the procedures and getting it to produce Python scripts has taken effort, and the automation remains unfinished.
That is the gap between an agent demo and a dependable assistant. A task can sound simple in a prompt, while the real work lies in explaining the steps, connecting the right tools, and handling failures. Local models do not remove that work. They change the privacy terms under which it can be done.
A tool, not a digital companion
The early verdict is neither that local AI is effortless nor that it is useless. Hermes handled some organizing and analysis tasks well, while the daily briefing proved fragile and benchmark automation is still in progress. Even a powerful model does not turn a computer into an assistant that can reliably do anything asked of it.
The tester treats Hermes as software, not a digital friend, and intends to remain cautious while trying more models and hardware. That seems like a sensible posture. The promise is not a machine that understands everything, but a tool that can take on selected work without sending its inputs elsewhere. The privacy may be the easy part. Deciding what to delegate, and checking that it worked, is still on the human.