Safety · 2 min read
An AI that can control your PC, safely: how approvals work
Letting an AI change your volume, open apps or run commands sounds risky. Here's how to keep it useful and safe: what runs straight away, what asks first, and what never runs as administrator.
An assistant that only talks is half an assistant. The useful part is "turn it down a bit", "open Spotify", "how much space is left on C:?" and having it actually happen. The worry is obvious: what stops it doing something you didn't want? Here's how layman draws the line.
Three levels
| What you ask | What happens |
|---|---|
| Simple, reversible things: volume, brightness, media, opening apps, settings pages, folders, websites, dark mode, a new email | Done straight away, shown as a card |
| Sleep, restart, shut down, sign out | Asks you to confirm first |
| Anything that needs a command | Shows you the exact command and waits for Run |
Locking your PC is the one power action that doesn't ask, since it can't lose anything.
Commands you approve
For a request with no built-in action, like "what's my IP address?" or "how many files are in my downloads?", the AI writes one PowerShell command (zsh on a Mac) and puts it on a card:
- Nothing runs until you press Run. Every command the AI writes, every time.
- Commands that delete or change things get a warning: "This command can delete or change things on your PC. Check it before running."
- Never as administrator. Commands run with your normal permissions, so anything that needs admin rights fails and says so.
- A time limit. A command is stopped after 60 seconds.
- An approval you ignore cancels itself after two minutes.
The commands are written by a model, so read them before you run them. A wrong command usually just returns nothing or an error, but the approval step is there for a reason.
Questions aren't orders
"How do I change my wallpaper?" gets an explanation, not a changed wallpaper. layman acts only when you tell it to do something.
Several steps at once
Type "open spotify and set volume to 30" and the bar shows Will run 2 steps, each with its current state. Drop a step with ✕, run them all with Enter, or save the pair as your own command to reuse by name.
Turning it off
Don't want actions at all? Switch off "Do things on this PC" in Dashboard → Settings, and layman will only ever answer in words.
Things to try
- "a bit quieter", "brightness 40", "next song"
- "open bluetooth settings", "open my downloads"
- "play lofi on youtube"
- "email sam about friday"
- "remind me at 4 to call mum"
Related
More guides
- How to run AI locally on Windows (no cloud, no account)A plain-English guide to running a ChatGPT-style AI on your own Windows PC: what you need, which model to start with, and the easiest way to set it up.
- Which local AI model can your PC run? A RAM and GPU guideMatch a local AI model to your computer's memory and graphics card. Sizes for Qwen3, gpt-oss, Gemma, Phi and Mistral, plus a simple rule for what runs fast.
- How to summarize a PDF with AI without uploading itSummarize contracts, reports and Word documents with AI that runs on your own computer, so confidential files never get uploaded to a chatbot.