A voice assistant on a desktop is one of the most privileged programs you can install. It holds the microphone, it can read what's on your screen, and it can execute commands. Any one of those is a serious grant. A tool that asks for all three should expect serious questions. Here are six, and why each one matters.
One: what listens while the assistant is idle? Soren’s current preview uses push-to-talk, so no wake-word process is listening while idle. The planned hands-free activation path uses an on-device wake-word model.
Two: when exactly does audio leave the machine — and does it leave at all? Listen for a precise answer. Vague answers (“we take privacy seriously”) usually mean “after some processing, it goes to us.” The precise answer for Soren: audio never leaves. Recording happens in a bounded window while the hotkey is held, and transcription happens on the Apple Neural Engine. Transcripts and relevant task context may be sent to the selected model provider for reasoning; depending on the task, that context can include tool results, UI text, or screenshots.
Three: what stands between a spoken sentence and a destructive action? Speech is a noisy input channel — you will be misheard eventually, and the tool should be built for that day. Ask whether destructive operations require explicit confirmation, and what “delete” physically does. In Soren, destructive actions stop at an on-screen approval prompt and wait for your yes, by voice or click; deletes are moves to the Trash, which Finder can undo; and saying “stop” cancels whatever is mid-flight.
Four: where do API keys live? If you bring your own model key, it should be in the macOS Keychain — encrypted, access-controlled, auditable — not in a plist or dotfile. That's where Soren keeps yours.
Five: can you see what it did? Every Soren action is logged as a readable trace — which tool ran, against which app, when. An assistant that acts on your machine without a reviewable record is asking for trust it hasn't earned.
Six: can you verify any of this yourself? This is the question that keeps vendors honest, because the first five can all be answered with confident fiction. Watch the process's network activity while the assistant idles. Hold the hotkey, make a request, and check when packets move. Trigger a delete and look in the Trash. Every claim above is checkable from your own machine in ten minutes.
If you run a fleet of Macs rather than one: all of the above still applies, plus disclosure. Voice tools in shared or monitored environments should be announced to the people in the room, and rolled out under your MDM policy like any other privileged software. The checklist isn't just for you — it's the same list a security review will run, so it's worth having the answers before the review does.