Walk through Offline Voice Control
7 of 11 screens are live in the demo. Everything else is visible but marked, with a reason — nothing is a dead link.
How to sign in
The demo has its own login because that is part of the product. These accounts are fictional, the passwords are deliberately fake, and each one sees a different slice of the platform — so it is worth signing in as more than one.
| Role | Password | What they can see | |
|---|---|---|---|
| Enrolled voice | [email protected] | Nova@Demo | Everything. Standing in for `python -m agent.main` — the product has no login. |
What to try first
- Run the clip called "Whisper handed silence". It returns a fluent sentence about subscribing to a channel, from a room where nobody spoke.
- Run "The right words, the wrong voice". Everything a text-only pipeline checks would have passed it.
- Run "start mkali" — it is held for a voice challenge. Pass one, then run it again.
- On the challenge screen, try the replay and the impostor. Each defeats exactly one of the two tests, which is why there are two.
- Read the Thresholds page for the two numbers the real config marks as guesses to measure yourself.
Every screen, and whether it is in the demo
Grouped the way the real platform groups them. A screen marked not in the demo is still visible in the demo's navigation — clicking it explains why and offers a walkthrough, rather than doing nothing.
Sign in
Stands in for `python -m agent.main`, and says so. The product has no login — it answers to an enrolled voiceprint.
Pipeline
The screen this demo exists for: twelve clips through eleven gates, each stage naming what it measured, what governs it and what it decided. Six of the clips should be refused.
Voice challenge
A phrase drawn fresh from sixteen words, checked for the words and for the voiceprint. Four scenarios — you, an impostor, a replay, and a noisy room — each failing a different way.
What it may do
The allowlists: twelve intents with their mechanism and gate, five whitelisted commands, two VMs, the Chrome profiles and the app aliases. A shell string is never built from audio.
What it finds
Applications scanned from the standard folders, settings panes from Apple’s URL scheme, files through Spotlight — so the agent never carries a stale table of what exists.
Thresholds
Every number the pipeline checks against, what moving it costs in both directions, and the two the real config marks as placeholders to measure on your own hardware.
Action log
logs/actions.log. Records refusals as well as actions, because a refusal you cannot explain afterwards is indistinguishable from a bug.
Live microphone
The real agent listens continuously, gated by voice activity detection and a wake word.
A browser could ask for the microphone, but the models are a 140 MB download and a second of CPU per clip. So the clips are fixtures, and their measured properties are real ones.Spoken replies
The agent answers through macOS `say`, using the Samantha voice at rate 200.
Browser speech synthesis would sound like a different product. What the agent would have said is shown as text instead, which is what the terminal shows too.Running the actions
Opening apps, starting VirtualBox machines, and running allowlisted commands over SSH.
There is no VirtualBox, no remote host and no macOS to script from a browser. The intents resolve fully and the log records what would have run, with its arguments.Enrolment
`enroll.py` records about thirty seconds of speech and writes a voiceprint, then reports a threshold to set.
It needs a microphone and the encoder model. Enrolment is a toggle here, and the challenge screen shows what the report is for: picking a threshold that neither locks you out nor lets your brother in.Good to know before you open it
- It runs entirely in your browser. There is no server, and nothing you type leaves the page.
- Your edits persist across a refresh, and Reset data in the demo bar puts the sample data back.
- A guided tour starts on your first visit and can be replayed from the demo bar.
- Dark by default, matching the real platform. There is a light/dark switch in the demo bar.
- It works on a phone as well as a desktop.