The software is free, and it stays free. What costs money is the machine it runs on and the AI inside it, and you pay both providers directly: a cloud server in your own account at Hetzner (about €6 to €21 a month on the Arm machines the wizard picks first, more where a location only sells Intel or AMD) or DigitalOcean (roughly US$24 to US$96), depending on the size you pick in the setup wizard, billed to you by them. These are indicative: the wizard shows your host's own current list price before anything is created. your own Claude subscription, billed to you by Anthropic nothing to Crads-AI for the software, ever Those facts are also the first screen of the wizard, before anything is created, so the bill is never a surprise. If you want a person to set it up with you or support you afterwards, that is a separate, optional service (Crads AI offers one, and so can anyone else who knows the software); it is never a software fee.

Questions, answered straight

Q: Do I need a Claude account?

Yes. Your assistant runs on Claude, on your Claude account, on your machine. That is deliberate: it keeps everyone else out of your conversations. You sign it in once during your first hour, and scheduled jobs run on that sign-in.

The questions people actually ask before and after getting a mineral, each answered in a few sentences with the deep page beside it.

Short answers, in the order people ask them. Every answer links the page that carries the detail, and nothing here says anything the deeper pages do not.

What actually is this?#

A private AI assistant that lives on a machine of your own (we call the machine a mineral), knows your world because it interviewed you, and works for you on a schedule: inbox sorted before you wake, a brief that knows what matters, follow-ups chased. What an AI EA actually is makes the full argument; what a mineral is explains the two words you will keep meeting.

What does it cost?#

The software is free, and it stays free. What costs money is the machine it runs on and the AI inside it, and you pay both providers directly:

a cloud server in your own account at Hetzner (about €6 to €21 a month on the Arm machines the wizard picks first, more where a location only sells Intel or AMD) or DigitalOcean (roughly US$24 to US$96), depending on the size you pick in the setup wizard, billed to you by them. These are indicative: the wizard shows your host's own current list price before anything is created.
your own Claude subscription, billed to you by Anthropic
nothing to Crads-AI for the software, ever

Those facts are also the first screen of the wizard, before anything is created, so the bill is never a surprise. If you want a person to set it up with you or support you afterwards, that is a separate, optional service (Crads AI offers one, and so can anyone else who knows the software); it is never a software fee.

Can you read my stuff?#

No, and under this design the answer is unusually short. Your mineral runs on your own hosting account, its only key is yours, and nothing about it routes through Crads AI: no account, no directory, no billing, no tunnel. There is nothing of yours for us to hold. The one way anyone else gets in is a support window you deliberately grant, which expires on its own clock. The privacy policy is the binding version. What does go to Anthropic, on your own Claude account, is laid out on self-hosted Claude: what runs where.

Do I need to be technical?#

Mostly no. If you can make an account at a hosting company and paste a token, the wizard does the rest, and get a Hetzner API token (or the DigitalOcean one) walks the one genuinely unfamiliar step. Answering the interview needs no technical skill at all. The deliberately technical route exists too, for those who want it: Claude Code on your mineral.

Do I need a Claude account?#

Yes. Your assistant runs on Claude, on your Claude account, on your machine. That is deliberate: it keeps everyone else out of your conversations. You sign it in once during your first hour, and scheduled jobs run on that sign-in.

Why doesn't installing the Claude Code app count as signing in?#

Because it signs in the wrong machine. The desktop app signs in your laptop, which then reaches into your mineral over SSH; scheduled runs happen with nobody at the keyboard, so they need the mineral's own sign-in. That is one claude run in the Terminal tab, once. Claude Code on your mineral explains why the two are different.

Where is my data kept?#

On the server you created, in the location you picked in the wizard, in your own Hetzner or DigitalOcean account, plus a backup in a private GitHub repository in your own account. The wizard offers whatever locations your host sells (DigitalOcean has Sydney; Hetzner does not), so where your data lives is your choice, and if you are in Australia and pick a European location, your data is stored overseas: the privacy policy says so plainly, because Australian privacy law expects it and you should know regardless.

Can it run without a server?#

Yes, though it isn't the path we recommend. In the app, choose Set up my own; the last option, after the hosting companies, is On this computer. It is free and makes a brain folder on your computer that Claude Code opens. The catch: no schedule, no phone, and it stops when your laptop sleeps. Scheduled jobs, the morning brief and Telegram need a server. When you use it, your requests still go to Claude, same as on a server.

You can change your mind either way. Move to a server on Your mineral builds a server and takes the folder's brain, connections and sign-ins onto it; Move to this computer does the reverse, and pauses the schedule and Telegram in the folder until you move back. See moving between your computer and a server.

It can also run on Claude's own cloud as part of your Claude plan, with jobs at most hourly and your files held by Anthropic; that one is set up with you in a guided setup.

Which servers is my data actually on?#

Two different machines get called "the server" and it is worth separating them.

Your mineral is the machine you rented in your own name. Your notes, your brain, your skills and your schedule live there and nowhere else, plus the backup in your own repository.

Anthropic's Claude servers are what your assistant sends a question to and gets an answer back from, on your own Claude account. They do the thinking. They are not where your world is kept: what goes over is the part of the question your assistant needs answered, not your filing cabinet.

Is it low-carbon?#

Built in Helsinki, the wizard's default, your server runs on Finland's grid: about 57 grams of CO2e per kilowatt-hour in 2025, roughly a ninth of Australia's average. Hetzner also holds hydropower certificates for its Finnish data centre. Two limits, plainly. If Helsinki is out of stock the wizard builds in Germany instead, where the grid is dirtier (Hetzner says it buys renewable electricity there). And the thinking happens at Anthropic, not on your server, so this makes the server cleaner, not the whole assistant. Pick DigitalOcean for Sydney and your server runs on the Australian grid.

What happens if Crads AI disappears?#

Less than you would think, which is the honest advantage of this shape. Nothing of yours runs through us at runtime: your machine keeps running, your backups keep landing in your own repository. What stops is new software and us. The software is open source, precisely so that even that gap can be picked up by someone else. What happens if Crads AI goes away answers this at length, starting by conceding it is the right question.

Can I leave? What do I keep?#

Everything, because you already hold everything. Your brain is readable files on a machine in your own account, and your backup is in your own GitHub repository. Stopping is deleting a server you own, in your own hosting console, whenever you choose. There is nobody to notify and nothing to claw back.

What is the interview, and do I have to do it in one go?#

The interview is the hour that turns a blank machine into an assistant that knows your world, and no, it resumes exactly where you stopped, days later if need be. Three twenty-minute sittings beat one exhausted hour. The interview explains the eight layers and what happens to your answers.

Can I rename my assistant?#

Yes. One name everywhere: your assistant's name is your mineral's name, and Rename (beside the name on the Your mineral page) changes that one thing. Your connections and your keys stay exactly as they are.

The brief got something wrong. Is that it?#

No, that is day two. Tell your assistant, in plain words, in the Terminal: what it got wrong and what matters instead. It writes that down and the next brief reads it. An assistant you correct gets sharper every week; your first week builds the habit.

How do I ask it for something properly?#

Plainly, the way you would brief a person, with the context it cannot have guessed. For anything that matters, use the CRIT method: say the context, give it a role, let it interview you before it answers, then name the task. Most of what looks like a bad answer is a thin brain, and grow the brain is the fix; the vocabulary these pages use is in the glossary.

Something looks broken. Where do I start?#

When something looks wrong lists the symptoms people actually hit, each with its cause and the page that fixes it. If a scheduled job went quiet, the answer is usually on Connections: jobs fail rather than pretend when a connection dies.

What happens when I disconnect a service?#

The credential is destroyed on your mineral, immediately. The permission you originally granted lives at the service's end and is yours to revoke there, in that account's own settings. The Secrets page shows what exists on your mineral without ever showing values, so you can check the result rather than trust it. Fix a broken connection covers the states a connection can be in.

Can I use it from my phone?#

Yes, and it is on already. Open the Claude app on your phone, signed in to the same Claude account as your mineral, go to the Code tab and pick your assistant. It is the whole assistant, with your brain, skills and connections, and the work runs on your mineral while your laptop is shut. Telegram stays useful for the morning brief and other results it sends you. Use it from your phone has the steps, what Anthropic keeps, and how to switch it off.

How do I use it from a second computer?#

A computer that already opens your mineral lets the new one in, directly, over SSH: no account, no service in the middle. Add another computer walks both halves; it takes about two minutes with both machines in front of you.

How is this different from just using ChatGPT?#

Ask a chat window "what should I focus on today" and you get advice-shaped filler, because it cannot know you. Your assistant answers from your actual goals, your actual inbox, your actual people, and it acts on a schedule instead of waiting to be asked. The longer answer, including what this is not, is what an AI EA actually is.

Who owns what I write? Do you train on it?#

You own your content, and it never reaches us, so there is nothing we could train on. Your assistant's answers to you run on your own Claude account, under your agreement with Anthropic. The terms say all of this in the binding form.

Do I keep using the chat window and the browser extension?#

No. Once your assistant is running, this replaces them. You do not need the browser extension and you can remove it, and you do not need to start work in a chat window any more.

The reason is not that they are bad, it is that they start from nothing every time. Your assistant already knows your goals, your people and your inbox, so work that used to mean pasting context in now just means asking. Your two interfaces from here are the Crads AI app, with the Terminal inside it, and, if you want a fuller window, the Claude Code app connected to your mineral.

Dictation software is the exception worth keeping: it types into any window, including this one, so if you already talk rather than type, carry on.

I have more than one Claude account. Does it matter which one I use?#

Use one, and use the same one every time. Your assistant signs in to Claude once and every scheduled job afterwards runs on that sign-in, so a second account is the most common reason a setup that looked finished stops behaving. If you have already signed in with the wrong one, type /logout in the Terminal, then /login, and sign in with the account you mean to keep. Scheduled jobs pick up the new sign-in from their next run.

How do I know what a skill actually does?#

Open it. Every skill on the Skills page carries a description of what it does and when it runs, and the ones your assistant came with are listed there alongside any you build. You can also turn a skill from something you ask for into something that simply happens on a schedule, and delete any you do not want. The Skills page walks through it.

Which model should I use, and why did I run out?#

Your assistant runs on your own Claude account, so the models you can pick and the amount you can use are set by your Claude plan, not by us. Stronger models think harder and use your allowance faster.

Usage runs in windows rather than as one monthly pot, so the useful habit is to check where you are rather than guess: type /usage in the Terminal and it tells you. If you are running out sooner than you expect, that is usually a strong model doing routine work, and asking your assistant to move the routine jobs down a level fixes it without you noticing a difference.

Are "pebble" and "mineral" standard industry words?#

No, they are ours. Nobody else uses them and you will not find them elsewhere, so if they read as jargon on first contact, that is fair. A mineral is the machine your assistant lives on. A pebble is one person's mineral, which is every mineral there is. What a mineral is is the longer version.