A Verge reporter is testing whether local AI can handle useful personal tasks without sending sensitive information to cloud services. Using Hermes Agent on an M5 Ultra Mac Studio, the reporter has sorted more than 400 Steam games and analysed financial records, while finding that setup and reliability still take real work.
Nous/Hermes Watch analysis
What happened
The reporter installed the open-source Hermes Agent and ran a 125-billion-parameter Qwen model locally on a Mac Studio with 256GB of unified memory. A Telegram bot made the agent controllable by phone. One early experiment, a daily briefing drawn from email and calendar data, repeatedly failed until the reporter realised the Mac could not be asleep when the scheduled job ran.
The more successful job was organising a Steam library by genre while preserving custom categories. It required a Steam API key, which the reporter says was revoked afterwards. The agent also analysed financial records and helped compare laptop specifications. A script to automate laptop benchmarks remains a work in progress. Read the Verge account.
Why it matters
This is a useful look at what a local agent can do today: organise a library, examine personal files and help with repetitive work, without the reporter uploading that sensitive information to a cloud service. It also shows the catch. Giving an agent access to useful tools means granting permissions, and even a straightforward scheduled task can fail for a very ordinary reason: the computer is asleep.
The account describes one person’s experiments, not a controlled test of Hermes or a guarantee that local processing makes every part of an agent private. The reporter says the briefing broke more than once, and the benchmark automation is unfinished.
Our read
The best case for local AI is not that it becomes a digital friend. It is that it can take on bounded, tedious jobs while keeping certain data on your own machine. This diary makes that possibility concrete, and makes the setup costs concrete too. Start with a task you can check, grant only the access it needs, and revoke credentials when they are no longer required. Glamorous? No. Sensible? Often.
What to watch
- Whether the reporter gets the laptop benchmark workflow working reliably.
- How useful the agent proves on smaller machines and less capable local models.
- What permissions and credentials each task needs, and whether they can be limited or revoked.
Discussion spark: Would you let a local AI agent access your email or financial files for genuinely useful tasks, or is the permission and setup burden still too high?
Sources and evidence
- Source update (11 October 2026, 11:00 UTC)
Nous/Hermes Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Nous Research.