Early accessMac only, and still being built.See exactly what works today
Docs

How Arkemis
actually works.

A short, honest guide to using it day to day — what it does locally, what leaves the machine, and what counts against your plan.

Summon
⌥⇧A
Modes
Agent · Transcribe · Dictate
Screen reading
On-device OCR
Confirm
Before destructive steps
01Guide

Getting started

Install the app, sign in through your browser, and press ⌥⇧A. The pill appears over whatever you are working in. Speak a request and release — Arkemis works out what to do from there.

There is no phrasing to learn. It is command-first rather than a chat window: you say the outcome you want, and it decides which tools it needs and whether it has to look at your screen.

02Guide

macOS permissions

Three permissions, each prompted on first launch:

  • Microphone & Speech — so it can hear you. Audio leaves only while you are speaking and is discarded once the text comes back.
  • Screen Recording — so it can read the active window when a task genuinely needs to see it.
  • Accessibility — for the global shortcut, for reading the structure of what is on screen, and for operating apps on your behalf.

If you skip one, the features that depend on it are unavailable until you grant it in System Settings → Privacy & Security.

03Guide

The pill and its three modes

One surface, three jobs. Tap the mode circle on the pill to switch, or let the request pick for you.

  • Agent — a request that may take steps. It reads the screen, searches, runs commands and drives apps, then reports what it did.
  • Transcribe — hold to talk; a hosted speech model returns your words in a fraction of a second.
  • Dictate — the same speech, typed straight into the field you are focused on, in any app.
04Guide

Screen understanding & computer use

When a task needs context, Arkemis reads the active window, your current selection, or the whole screen using on-device OCR and the Accessibility APIs. That is what lets “summarize this” or “reply to this” work without you copying anything.

When a task needs actions, it can click, type, move the pointer, launch or switch apps, and run commands. Anything destructive or irreversible stops for a one-tap confirmation first. Being agentic is reserved for work that needs it — most requests are single-step.

05Guide

How usage is counted

Two meters, so one kind of use never eats the other’s budget. Quick local commands are not metered at all.

A spoken instruction to the assistant is not a dictation — it is part of the AI Action it asks for. The live numbers for your plan are on the pricing page, read from the same record the backend uses to enforce them.

What counts as one unit
AI action
One complete hosted request
Agent task
One action, steps included
Research task
One action
Dictation
One spoken utterance
Local command
Unmetered
06Guide

Account, plan and billing

Authentication runs through Clerk. Pro is a recurring monthly subscription through PayPal; approval is verified server-side and webhooks keep your plan in sync. Cancel any time in your account settings and Pro stays active until the end of the billing period.

Your plan is resolved from the shared database that the backend and the Mac app both read, so the app and the website never disagree about what you are on.

07Guide

Desktop sign-in handoff

The Mac app never asks for your password. It opens your browser, you sign in with Clerk, and the browser hands a single-use code back to the app over a custom URL scheme.

Callback
URL scheme
recall://auth?code=<single-use-code>