How Zeph works
Ships today

Workflows

A trigger, a few steps, something happens. Zeph's automation primitive — closer to Shortcuts than to a chatbot.

A workflow is an automation: a trigger, a short chain of steps, and an effect on your machine or your device. Hold a button, speak, and the text lands at your cursor — that is a workflow, and it is the shape most of them take.

Each one has a name, an icon and a colour, and any number of triggers can point at the same workflow. You build it in the companion as a list: add a step, expand the card to configure it, drag it into place. Mac Shortcuts is the reference the editor was drawn against.

What a workflow is made of

Steps are typed, and they only connect where the types line up. The Dictation starter is three steps and a fan-out at the end — the transcript goes to your cursor and to the assistant panel together.

The Dictation starter

Capture audioListen to the device microphone.
audio
Adjust audioLevel and clean up the audio.
audio
TranscribeAny speech provider you have configured.
text
All at once
Type at cursorInsert the text at the caret.
Notify assistant panelPost the transcript to the panel.
The words on the arrows are the types. Audio goes into a transcriber; text comes out. A step that does not match cannot be connected. The last step feeds every terminal in the group at the same time — in the editor, that group is called 'also send to'.

Dictation ships with review turned on: the transcript is parked for you to read, and typing happens when you accept it. The other setting is fire-and-forget — the result goes straight out.

The full vocabulary today:

StepWhat it does
Capture audioListen to the device microphone.
Device eventReact to an event from the device.
Adjust audioLevel and clean up the audio.
TranscribeTurn speech into text. Any speech provider you have configured.
Ask agentAsk the assistant and use its reply. Optional — most workflows never touch it.
SpeakSynthesize speech from text.
Press keysPress a key combination on this Mac.
Device actionTrigger an action on the device.
Record to fileSave the audio to a file.
Type at cursorInsert the text at the caret.
Copy to clipboardPut the text on the clipboard.
Notify assistant panelPost to the assistant panel.
Play on devicePlay audio on the device.
Virtual microphoneFeed the system virtual mic.

The three text destinations — cursor, clipboard, assistant panel — are the ones you can stack in the fan-out group. One transcript, up to three landings.

A fresh install has no workflows and no bindings. Onboarding writes your first one, and three starters are a click each: Dictate to cursor, Voice Assistant, Meeting Recorder. Each arrives as a real, editable workflow.

Where the assistant fits

It takes text and returns text, so it drops into the middle of the chain without disturbing either end:

The same chain, with the assistant in it

TranscribeSpeech to text.
text
Ask agentSend the text to a model and use its reply.
Optional
text
Type at cursorInsert the text at the caret.
Text in, text out. Leave the step out and the transcript is typed. Put it in and the model's reply is typed instead. Dictation and Meeting Recorder both leave it out — neither contains a model.

The assistant is one step in the list.

The companion runs with no assistant configured, and workflows built without one keep working — Dictation and Meeting Recorder contain no model at all.

A workflow is where AI becomes reachable.

What fires one

Four things, today:

  • A button on your device. Main or aux; tap, double-tap, or hold.
  • A keyboard shortcut on your Mac. Press to open, release to finish — hold-to-talk without touching the device.
  • Incoming device audio — audio arriving from a device you have routed.
  • Manually, from the companion.

Where each trigger applies — and how the same workflow can sit on a button and a keyboard shortcut at once — is the job of a binding. The binding is also where a running workflow shows its state.

One split is worth knowing here. A keyboard-shortcut workflow fires on your Mac and puts nothing on the device. The device is not involved: your Mac hears the shortcut, runs the steps, and presses the keys. So a workflow that only presses keys never takes a slot on the device's screen.

There is no scheduling. No time-of-day, no location, no app-launch, no webhook. Do not plan around scheduled triggers yet.

Why scheduling isn't here yet

Scheduled and agent-initiated triggers were taken out rather than shipped half-built: they existed in the design but were never connected to anything that would fire them, and a binding that silently does nothing is worse than no binding at all. They are tracked to return.

Where the Shortcuts comparison stops

The comparison is the one the editor was drawn against, so be clear about the ceiling: the action vocabulary is small and audio-centric. There is no web-request step and no scripting step. The steps in the table above are what a workflow has to work with today.

Where this is going

The vocabulary grows, and two moves past it are already drawn.

A workflow will be able to trigger another workflow, so a short chain can hand a result to a longer one instead of repeating its steps. That composition is a direction, not something you can build yet.

And a control on a glance will be able to fire a workflow when you tap it — a shortcut on the device's screen, the way a complication sits on a watch face. That one is designed, and waiting on a device capability to land.

On this page