Security
You add one line of code to your website. Here is exactly what that line does, what it never does, and what we're still improving.
The line you add
<script type="module" src="https://tinycube.cubemto.com/t/your-site-id.js"></script>
type="module"means the script loads after your page and keeps its code separate from yours. It never slows the page down while it loads.your-site-idis a random ten-character name for your site in TinyCube. It isn't a password: it only lets the script fetch your public search data.- The address loads the assistant for your site, and nothing else.
What the script does, step by step
- Builds the assistant: the button and its window, sealed off from your page.
- Fetches your site's search data when the browser is idle, and keeps it in the browser's cache.
- Checks what the device can handle and, once the visitor uses the assistant, loads the search model.
- Answers: searches for the visitor's question on their device and shows the matching text from your site. This is also where the visitor is asked before a question is passed on.
- Helps AI agents: adds three links to the page header and offers two search tools.
What it loads, and from where
Your visitor's browser only ever contacts TinyCube. Nothing is loaded from any other company's address.
| What | When |
|---|---|
| The assistant's code (four small files) | When a page opens |
Your site's knowledge base, as one .tinycube package (about 1–2 MB) | When the browser is idle, or when the assistant is opened |
| A code library that runs the model in the browser | Only after a visitor opens the assistant or asks a question |
The search model, as one .tinycube package (about 34 MB, kept on the device afterwards) | Only after a visitor opens the assistant, on devices with enough memory, and never when data saver is on |
The model and the library are stored with TinyCube on Cloudflare, as part of TinyCube itself. When a visitor's browser needs them, Cloudflare sends them directly: no other host is contacted, by the browser or by us. We take a copy of the model from its publisher once, when we build a new version of TinyCube, not when your visitors use it.
About .tinycube packages and keys. Your knowledge base and the model are handed out only in TinyCube's own package format, and every package is locked. Your knowledge base is locked with a key of its own, and that key is given only to pages on the addresses you added for your site in the portal. A copy of your script tag on someone else's website gets no key, and the assistant doesn't appear there. The loose files are not served at all.
What this does and doesn't do: it keeps your knowledge base and the model from being used on other websites or dropped into another tool. It is not a vault. A page of your own site has to unlock the package to answer visitors, so a determined person using your site can still get at what it contains, and your public pages are public anyway. The model includes open-source work used under its licence; the notices travel inside the package.
What it sends
Visitors' questions are searched on their own device. They are not sent to TinyCube, to an AI company, or to anyone else.
There is one exception, and it needs two yeses. If you switch on "let visitors pass on unanswered questions" in your portal, a visitor whose question had no answer is asked whether to send it to you. Only if they agree, the question and the address of the page they were on are sent to TinyCube and shown in your portal. No name, email address or other details are collected.
A second option works the same way: if you let visitors leave their email address for a reply, a visitor can type their address and press Send. Their question and address then reach you, and the address is deleted when you mark the enquiry done.
The script does no tracking. It does keep a tally for you: each question adds one to a count of questions asked, answered and unanswered for your site, which you see in the portal. The count carries nothing about the visitor and nothing of what they typed.
What it stores on the visitor's device
- The search data and the model, in the browser's cache, so they aren't downloaded again.
- Two small notes in the browser's storage: whether the visitor left the assistant open, and whether a search level failed on this device so it isn't tried again.
- The current conversation, for that browser tab only, so it is still there when a suggested link opens another page. It is gone when the tab closes.
It sets no cookies and stores nothing that identifies a person.
What it changes on your page
- It adds the assistant: one element at the end of the page, sealed off from your page's styles so neither affects the other. It sits above everything else, and moves up if you already have a button in the same corner.
- It adds three links to the page's header for AI tools: your MCP server, your
llms.txt, and a Markdown version of the page. - In browsers that support it, it offers AI agents two tools that search your site's knowledge on the device. The tools can only read what is already public in your search data.
What it never does
- It doesn't read what visitors type into your forms, or their passwords.
- It doesn't read or set cookies.
- It doesn't change your content, your links or your prices.
- It doesn't write its own answers. Every answer is text from your pages, your documents, or an answer you wrote in the portal.
Your own styling
The CSS you add in the portal is applied only inside the assistant. We keep it to plain styling: anything that would load a file from another address is removed.
Updates
The script is served from TinyCube, so improvements reach your site without you changing the line. That also means you are trusting us with what the script does, as with any tool installed this way. Browsers keep a copy for five minutes. We plan to offer a fixed version you can pin, so that the script only changes when you choose.
Your account and your data
- Your password is never stored. We keep a scrambled form of it that can't be turned back into the password.
- You stay signed in through a cookie that scripts can't read, for seven days. Six wrong passwords lock sign-in for fifteen minutes.
- Sign-ups, sign-ins, reset requests and preview builds are limited per network, to slow down abuse.
- Payments are handled by Stripe on Stripe's own pages. TinyCube never sees or stores card details; we keep only whether your plan is active.
- Previews of other sites are shown in a sealed frame of the browser: a previewed page cannot read your TinyCube sign-in or act as you.
- Only your account can see or change your sites, documents, questions and answers.
- Your site's text and search data are stored with Cloudflare. Free previews are deleted after one hour.
- Documents you attach are read in your own browser. Only their text is sent to us, not the file.
How we read your website
We read only public pages, follow your robots.txt, read slowly, and identify ourselves as TinyCubePreview. To keep us out, add that name to your robots.txt.
What we're still improving
- A fixed, pinnable version of the script.
- Switches in the portal for the AI links and for the model download.
- Checking that a new account's email address is real, and sending password-reset links by email. Until the email part is switched on, write to us from your account's address if you forget your password.
- Meaning search in languages other than English. The assistant speaks English, Dutch, German, French and Spanish, but outside English it finds answers by matching words, not by meaning.
- TinyCube is in early access and has not had an independent security review.
Found a problem, or have a question about any of this? Write to cubemto@gmail.com.