Running an AI model on your own computer can make it easier to try tasks that involve private information. But local AI still asks you to figure out what to delegate, set up access, and troubleshoot when an automation fails. One experiment with Hermes Agent and a Qwen model shows both the appeal and the day-to-day friction.
Privacy makes experimentation feel different
The trial began with a hesitation many people may recognize: sending personal data to a cloud service. A locally run model offers another way to explore AI, with the software and model running on a computer rather than relying on a chatbot subscription for each task.
That distinction mattered for work involving financial records and information under embargo. The experiment used Hermes to analyze financial data and create a laptop specification comparison spreadsheet. Keeping those materials on the computer made it possible to use AI for tasks the writer would not entrust to a cloud service.
Local use does not make every task worthwhile, though. The question is still whether the time and expense of suitable hardware are justified by what the system can do. The early examples suggest that private, practical jobs may be a more convincing starting point than expecting an all-purpose assistant.
Getting started is simple; choosing a task is harder
Hermes Agent is an open-source, self-hosted desktop app available for macOS, Windows, and Linux. It can be used for free with local language models. On an M5 Ultra Mac Studio with 256GB of unified memory, the writer selected Qwen 3.8 Flash Next, a 125-billion parameter model around 105GB in size.
The app's onboarding and model picker made installation manageable, and a Telegram bot made it possible to control the setup by phone. Yet the first obstacle came right after setup: an empty text box and no obvious answer to what the system should do.
A daily morning briefing provided a low-stakes first experiment. Hermes scanned email and a calendar, highlighted items needing attention, and added a short weather report. The first attempts failed because macOS was asleep when the scheduled job ran at 7:30am. Once the computer stayed awake, the briefing worked, although its usefulness still depended on finding more relevant information for it to gather.
Repetitive organization is a natural fit
A more satisfying job was sorting a Steam library of over 400 games. Steam required games to be categorized manually, so the writer asked Hermes to reorganize the collection. The agent could inspect most games through the installed Steam client and proposed ways to arrange them.
The chosen approach grouped games by genre while preserving personal categories such as favorites, co-op titles, party games, and games to play with his wife. For this task, Hermes needed a Steam web API key. After the sorting was complete, the key was revoked. The result was a library that was easier to browse, with new purchases left for the writer to categorize later.
This example also shows the tradeoff behind an agent that can act on a computer: it needs permission to reach the relevant tools and information. A limited task can be useful, but granting access is part of the setup and should match what the job requires.
Automation still needs supervision
The more ambitious project was automating laptop benchmark tests. Running a dozen or so tests three times each to calculate an average takes time, so the writer began guiding Hermes through the procedures and asking it to produce Python scripts. That work was still underway.
The broader lesson is that local AI can take on useful busywork without becoming a dependable assistant for everything. Even the morning briefing broke multiple times, and the benchmark automation required careful instructions. Starting with a specific, repetitive job can make the value easier to judge, while keeping expectations grounded in the work still needed to make the system reliable.
For people wary of uploading sensitive records, local AI may open up tasks they otherwise would avoid. The experiments show tangible uses in organizing, summarizing, and handling data, alongside the need to choose the right task, grant access thoughtfully, and repair automations when they fail.