First run
The setup wizard: connect a device, set up audio, and pick what you want Zeph for.
Open the companion for the first time and it hands you a full-window wizard. Two minutes, give or take. It finishes when your device has actually done something — a real transcript, a real tool call. That is what marks a track complete.
The five screens
Four of them are the same for everyone. The fifth is the one you chose.
Welcome
One card, one Continue. Continuing is also where you accept the Terms and the Privacy Policy, so that lands before any of the rest can be skipped.
Get your device
Press a button to wake your device and it appears under Detected devices, over Bluetooth or USB. Three ways forward, and all three are real exits:
- Connect a device we found.
- My device isn't listed — the flashing flow runs inside the wizard, in place, and hands you back to it when the board comes up. Flashing over USB during setup erases first, so a device you re-onboard starts factory-fresh and unpaired.
- I don't have it yet — the device steps wait for you and setup carries on.
Bluetooth off? The gate appears right there, on this screen, with the fix — or plug in over USB instead.
Set up Zeph audio
Optional, and it asks macOS for your password — once. It installs the audio driver, which makes your device appear on the Mac as two devices: a wireless microphone and a wireless speaker.
If the driver is already installed, this screen never renders. It passes silently and the walk moves on.
Choose what to do
The fork. Pick as many as you like; Start setup lights up at one.
Set up and try
Your tracks, back to back. Then the diagnostics question — one toggle, on by default, asked here rather than up front, and changeable in Settings forever after. Then the finish screen, which shows you the device face you are about to get and offers the track you did not pick.
What you want Zeph for
Three answers, and the checkboxes are yours to combine.
| Track | The steps it walks | What ends it |
|---|---|---|
| Transcribe at cursor | Audio driver · Accessibility · Speech provider · Try it | A real transcript arrives. |
| Use as a microphone | Audio driver · Accessibility · Pick the app · Try it | You confirm it works. |
| Use with agents | Connect · Approve · Install the skill · Try it | Your agent makes its first call. |
A shared step runs once. The two audio tracks both need the driver and both need Accessibility, and picking both does not ask you twice — the walk is the union of your tracks, de-duplicated. The driver screen on the spine and the driver gate inside those tracks are the same step: install it up front and it is already gone from the rest of the walk. Accessibility behaves the same way. Grant it once and the screen stops appearing.
The speech provider step is the real one, not a rehearsal: on-device models download to your Mac and stay there, and a cloud key goes into your keychain and is verified against the provider before the step will let you past it. Whichever you set becomes the system speech default. See Choose a speech provider.
The agents track mints a genuine access token when you copy the connect command — it appears in the copied snippet and nowhere else — scoped to device control and app install. You approve the client the first time it connects, install the zeph-device skill into it, and the track completes when the agent actually calls a tool. See Connect Claude Code.
The microphone track asks which app you will talk into and sets its push-to-talk key: Wispr Flow holds Fn; Zoom, Meet and Slack just want Zeph selected as the input device. Its "Try it" is the one win you confirm yourself, because your voice is streaming into somebody else's app and the companion cannot see inside it. The live meter is the proof. See Use as a microphone.
What setup writes to your device
It reads what you actually configured — provider, assistant, driver, app — and writes the smallest thing that follows from it.
Seeding, on connect
Your device keeps its own name. It names itself when it first boots, and setup leaves that name alone.
Apps come later, and from you. Your device arrives with a working home screen, and you add apps to it from the Library when you want them — see Install an app.
The hold it sets up for you
Main·hold — hold the main button on the home glance — is the gesture every track is built around. The companion fills that slot from what you actually have.
How the hold gets filled
Transcribe at cursor, with a speech provider set?
- YesDictation, to your cursor
- NoUse with agents, with an assistant provider set?
Use with agents, with an assistant provider set?
- YesVoice assistant, to the speaker
- NoIs the Zeph audio driver installed?
Is the Zeph audio driver installed?
- YesAudio, to the virtual microphone
- NoThe slot waits for you.
- Dictation, to your cursorThrough the provider you picked, a minute ago, on the speech step.
- Voice assistant, to the speakerHold and ask; the reply comes back through the device.
- Audio, to the virtual microphoneThe driver is there, so the mic route works.
- The slot waits for you.Open, and one click from Buttons in the companion.
Set none of those up and the hold waits for you. The glance says so — No hold action yet, with a link straight to Buttons — and you fill it there whenever you like. See Bind your buttons.
Pick Use as a microphone and the hold is compound — it routes the audio and holds your app's push-to-talk key for as long as you hold the button. One gesture, both halves.
Skip, resume, re-run
Skip — I'll explore sits pinned to the rail on every screen, and it drops you into the app.
Come back — from the sidebar card, or from the getting-started hub — and you land on the first screen you have not finished, not back at Welcome. What you already did stays done.
Settings → General → Re-run setup starts the wizard again from the top. It clears the wizard's progress and nothing else: your permissions, your providers and your device stay exactly as they are.