Privacy policy

What is stored, where it runs, what leaves your mineral, and why almost none of this page is about us any more.

Version 0.21, effective 2026-10-17. Samuel Davis trading as Crads AI.

Who this is#

Crads AI is operated by Samuel Davis, trading as Crads AI, ABN 26 929 349 775, in Coogee, New South Wales, Australia. Contact: cradsdavis@gmail.com.

This is version 0.21, effective 17 October 2026. The version and date at the top of this page change whenever the text does, and prior versions stay on the record.

What changed in each version is listed under Changes at the end.

The short version#

Your mineral is a machine in your own hosting account, in a location you chose, opened only by your own keys, running your assistant on your own Claude account. Your brain lives on it as ordinary files. Nobody at Crads AI reads your notes, and under this design the sentence is structural rather than behavioural: nothing of yours routes through us, and we hold no account, record or credential about your mineral. It is your data, on your machine, end to end.

The rest of this page is that sentence, in detail.

What we hold about you#

For a person who downloads the software and runs a mineral: nothing. There is no account to create, no registration, and no telemetry or analytics in the app. The few things the app remembers (which mineral you opened last, which side-menu groups you folded) live in your own computer's storage and are never sent anywhere. The app also keeps the SSH key it made for you, and a connection entry for each mineral, in your computer's SSH folder. Your hosting token is never saved; it is held only while the server is being built.

If you correspond with us (email, a support engagement, a paid service), we hold that correspondence and the ordinary business records of it, together with any roadmap call recordings you agreed to and, if you are on it, the essay mailing list below.

If you buy a paid service. Stripe takes the payment; we never see your card number. We receive your name, email, what you bought and the payment reference. Paying also creates a client account on crads-ai.com holding your name, email, Stripe customer reference, what you bought and how many sessions you have used, so you can sign in and book the rest. Bookings, paid or free, are made through Cal.com, which holds your name, email and booking time. Calls run on Google Meet.

Where you heard about us. If you answer "How did you hear about this?" at checkout, or the link you booked from carried a tag saying where it was posted (for example utm_source=linkedin), we keep that answer and that tag with your booking, on the Stripe payment and the Cal.com booking. Answering is optional.

If a partner sent you. Some people we work with have a referral link (crads-ai.com/r/ and their name), and we pay them A$100 if someone they sent buys a Guided Setup. If you arrived through one, or name someone in the "Who sent you?" box at checkout, we keep that partner's name with your booking: on the Stripe payment, on the Cal.com booking, and against your email on crads-ai.com for 90 days. If you buy a Guided Setup, we keep a record of the referral (your name, email and booking reference, and the partner's name) so we can pay them. We tell the partner your name, so they know what they are being paid for, and nothing else.

How we handle it. Our email, calendar and documents are in Google Workspace. We use an AI assistant, running on Claude under our own Anthropic account on a server we rent in Sydney, to read and draft our email, keep notes on the people we work with, and help write roadmap reports. Nothing it drafts is sent without a person checking it first.

If we wrote to you first. If we emailed you before you had written to us, we found your name, business and work email in public sources such as your website. We keep them, with a note of what we sent, so we write to you at most twice. Ask and we delete them.

Roadmap calls. If you book an AI problem-solving roadmap, we ask at the start of the call whether we can record it with an AI note-taker. If you say yes, the call is recorded and transcribed by the note-taking service (Otter.ai, acting for us), and we use the transcript to write your report. If you say no, nothing is recorded and we take our own notes. Setup sessions (the walkthrough and guided setup) are not recorded. You can ask for a recording or transcript of your call to be deleted: email cradsdavis@gmail.com.

The essay mailing list. If you are on it, we hold your email address, the page you signed up from, when you joined, which of the three ways below you joined by, and a random code that makes your unsubscribe link work. You join by signing up on the site and clicking the confirmation link we email you; by ticking the box when you book a discovery call; or, if you are already a client or have had a roadmap call with us, because of that existing relationship. The emails come from sam@crads-ai.com and are sent through Resend, an email delivery service. The list is kept in the website's database, run by Redis Cloud on Amazon Web Services servers in the United States. Every email has a one-click unsubscribe link. When you unsubscribe we keep your address marked as unsubscribed, only so that you are never added back; to have it removed completely, email cradsdavis@gmail.com. To stop the sign-up form being abused, the site counts sign-ups from each network address for 15 minutes, and that count then expires.

Visiting crads-ai.com#

The website, these docs included, counts visits with Vercel Web Analytics, run by Vercel, the company in the United States that hosts the site. For each page view it records the page address (including any tag on the link saying where it was posted, such as utm_source=linkedin), the site that linked you here, your approximate location (country, region and city), your browser, operating system and device type, and the time. It also counts two kinds of click: a link to book a call, and a link to download the app. Each records only the page you were on and the page the link pointed to. On the confirmation page after a booking, the payment reference in the page's address is removed before anything is recorded.

Web Analytics sets no cookies and stores nothing in your browser. The site sets two cookies: the sign-in cookie for the client account area, only if you sign in, and a referral cookie, crads_ref, only if you arrive through a partner's referral link. The referral cookie holds the partner's name and the date you arrived, nothing about you, and lasts 90 days. The app download itself is served from GitHub. Vercel tells one visit from another with a hash made from the request, which it discards after 24 hours, and it does not tie a visit to you or to your IP address, or collect anything that follows you to other websites. What we see is totals: how many people read which page, and where they came from. Nothing in it identifies you. If you block the script, the site works exactly the same.

The app, and your mineral, send nothing of this kind.

Where your mineral runs#

On a server you created in your own hosting account, at Hetzner or DigitalOcean, in whichever of that company's locations you picked in the setup wizard. Your contract for that machine is with them, not with us, and we cannot see, reach or bill it.

If you are in Australia and picked a location outside Australia, your data is stored overseas. We say that plainly because Australian privacy law expects disclosure of overseas storage, and because you should know it regardless: it is your choice, made in the wizard, changeable by rebuilding elsewhere.

If your mineral is part of a team#

If a business set you up through Crads for Teams, your server sits in their hosting account and your repository in their GitHub, and the owner holds a key that opens your server. Your personal/ folder and pages marked private are never published to the company brain, but whoever administers the servers can read them. Crads AI is not in that relationship: the team owner decides what happens to that data, and questions about it go to them.

What lives on your mineral, and stays there#

Your brain. The pages your assistant writes and reads: who you are, your goals, your people, your business, your voice.
Your credentials. Your Claude sign-in, your GitHub sign-in, your connected service tokens, your Telegram bot token if you use one.
Your run history. What your assistant did and when.

None of this is copied to us. There is no central store of brains, because there is no central store: one person, one machine.

What leaves your mineral#

Only what you set up, and each goes to an account of your own:

What you say to your assistant, to Claude. Every conversation and scheduled run sends your request, and the pages and connected data your assistant reads to answer it, to Anthropic under your own Claude account.
Phone sessions, to Claude. On a server, Remote Control is on unless you switch it off. It keeps an outbound connection to Anthropic so the Claude app on your phone, or claude.ai/code, can work with your assistant. While one of those sessions is connected, Anthropic stores its transcript (your messages, the assistant's replies and what it did) under your own Claude account, as with your other Claude conversations. The work itself runs on your mineral, and your files and connections stay there. Switch it off on the phone row in Connections; use it from your phone has the detail.
A backup, to a repository you own. A pebble pushes its brain nightly to a private GitHub repository in your own account, as readable files, and an encrypted snapshot of its settings and credentials to the same repository as ciphertext, AES-256, with a passphrase generated on your box. We never hold the passphrase and the repository is not ours. One property of git worth stating: the repository keeps history, so earlier snapshots persist in history you own until you prune it. GitHub stores repositories in the United States.
Things you asked it to send. Skill output you tick "Send me the result" on is delivered to your own Telegram bot, transiting Telegram's servers under your agreement with Telegram. Anything your assistant does through a service you connected (a draft in your Gmail, an event in your calendar) happens at that service, on your account, because you set it up and asked for it.
Update checks. Your mineral downloads its nightly update from GitHub's container registry, and the desktop app checks GitHub for a new version every few hours. GitHub sees the address these come from.

Services you connect#

Google (Gmail, Calendar, Drive). You create your own Google OAuth client and it stays on your mineral. Your Google key never touches Crads AI infrastructure, and we could not read your mail if we wanted to.
Everything else in the connections catalogue signs in on your mineral, with the token stored on your mineral.
Anthropic. Your assistant runs on your Claude account, under your agreement with Anthropic. What you send your assistant, you are sending through their service on your own account.

Because each of these is your account at that provider, connected by you, they are your providers rather than our sub-processors: we are not in any of those relationships.

When anyone can reach your mineral#

Three ways, all bounded.

Software updates. A pebble restarts nightly onto the latest published software. That updates the machinery, not your files, and you can switch it off. An update is a restart onto a published image; it is not a door, and nothing about it can read your machine.

Anyone signed in to your Claude account. With Remote Control on, a phone or browser signed in to your Claude account can start a session on your mineral with the same reach as you at the desktop: your brain, your skills, your connected services. So your Claude sign-in now guards your mineral as well as your chats. Keep the email or Google account you sign in to Claude with secure, and consider Claude's "Require trusted devices" setting (in beta), which asks each new device to confirm with a passkey or your fingerprint or face. Switch Remote Control off and this way in closes.

Support access, only if you grant it. If something is broken and you want help on the machine itself, you grant access explicitly. The grant expires on a clock enforced by the mineral's own door, and its lifecycle (that it was enabled, when, until when, and any revoke) is recorded on your own mineral where you can read it. What is not captured is a per-command transcript of the session itself; that is a real limit and we would rather name it than imply a log that does not exist. Nobody can add themselves to your mineral without you granting it, and there is no standing key: a mineral made by the wizard is born holding only your own. Remote Control is not a key: it is your Claude account.

What a support session can see#

Worth being exact about, because "someone can look at your mineral" and "someone can read your saved passwords" are different sentences and you should know which one you are agreeing to.

Your mineral stores secrets in two tiers:

Available to the box (the default for anything a scheduled job needs at 3am). Stored as a value on your disk. Encrypting it there would be theatre, because the box would have to keep the key on the same disk to use it at 3am. Anyone with access to the server can read these, and that includes a support session you have granted.
Sealed to your devices. Stored as an envelope only your own enrolled machines can open. There is deliberately no code on your mineral that can decrypt these, so a support session cannot read them, and neither can a scheduled job.

The Secrets page in the app lists every entry grouped by what it is for. The rule that decides the tier is the one above: anything a scheduled job needs while you are asleep is available to the box.

The same is true of your brain during a support session. A granted session is server access, and server access can read files. The honest statement is that the protection during an active grant is the grant record and the clock, not an inability: the mineral records that the window was open and when it closed, not each thing done inside it. If a credential would be damaging in someone else's hands and nothing scheduled needs it, seal it; if you change your mind about a grant, revoke it, and it ends immediately.

Taking your data, and deletion#

Your brain is a folder of readable files in a git repository on a machine you hold, mirrored to a repository in your own account. There is no export step, because there is nothing to export from: it is already yours, in a portable format, everywhere it exists.

Deleting your mineral is deleting your server, in your own hosting console, which destroys the machine and its disk under your contract with your host. Nothing needs our involvement and we hold no record of the machine to delete at our end. Copies you hold, including your backup repository, remain yours.

Breaches#

Given what is stored where, a breach of Crads AI could not expose your mineral's content, because we do not hold any. If something we do hold about you (correspondence, service records) is exposed, we will tell you what happened, what was exposed, and what to do, without waiting to have a complete picture first.

Complaints#

Write to cradsdavis@gmail.com. If you are in Australia and unsatisfied, you can complain to the Office of the Australian Information Commissioner.

Changes#

Changed in 0.21: "What we hold about you" says that the optional "How did you hear about this?" answer at checkout, and any tag on the booking link saying where it was posted, are kept with your booking; "Visiting crads-ai.com" says a page address includes any such tag.

Changed in 0.20: partner referrals. "What we hold about you" says what we keep when a partner sent you, and "Visiting crads-ai.com" discloses the referral cookie. The version line under "Who this is" now matches the top of the page.

Changed in 0.19: Remote Control (your assistant from the Claude app on your phone) is on by default on server minerals; "What leaves your mineral" says what it sends to Anthropic, and "When anyone can reach your mineral" says who it lets in.

Changed in 0.18: a section for minerals set up through Crads for Teams; "What we hold about you" now covers paid-service client accounts, Stripe, Cal.com, the business's own email and AI assistant, and people we wrote to first, and no longer calls itself the whole list; "What leaves your mineral" now names what you say to Claude and the update checks; the site's sign-in cookie is disclosed; this change list moved to the end of the page.

Changed in 0.10: the self-host redesign, which rewrote most of this page by deleting its subject matter. Minerals are now created by their owners, on their own hosting accounts, with their own keys; Crads AI operates no accounts, no registration directory for them, no billing, and no tunnels, and provisions nothing. The sections of versions 0.1 to 0.9 covering the registration directory and its email hashes, the community status heartbeat and its sharing switches, the setup key a community-built mineral was born with, EU-only hosting, and the sub-processor table describing our vendors are gone because the machinery they disclosed is gone. What survives is the support-access disclosure and everything about the machine itself, which was always yours. Changed in 0.11: the community model (a shared repository your mineral pulled read-only) was removed from the software on 2026-09-09, so the one sentence about it is gone; nothing else changed. Changed in 0.13: roadmap calls, recorded only with your OK, added under "What we hold about you", and the version line above corrected. Changed in 0.14: the crads-ai.com website, these docs included, now counts page views with Vercel Web Analytics, so the sentence saying the docs site had no analytics is gone and "Visiting crads-ai.com" below says exactly what is recorded. Nothing about the app or your mineral changed. Changed in 0.15: the website also counts two kinds of click (into booking, and into the app download), and the booking confirmation page no longer records the payment reference in its address; both are described under "Visiting crads-ai.com". Changed in 0.16: the essay mailing list, which started on 11 October 2026, added under "What we hold about you". Changed in 0.17: where that list is stored, now named (the United States) instead of "may be outside Australia".