Skip to content

Safe designs

A design can change how the app looks, but it cannot send your data away, fake a screen, or delete or share anything. Anything someone else made that can do something starts switched off.

The two doors

Almost every attack needs a way out (to send your data to someone) or a way in (to act as you). Designs get neither.

Settings, CSS and HTML have no way out and no way in. Scripts and HTML blocks run in sealed boxes and get only the doors you approve. Protected screens are walled off from every design.Settings, CSS, HTMLchecked when savedcannot run codeno web addressescannot reach the internetScripts, HTML blocksrun only in sealed boxesno screen, keys, filesa website only ifyou approve itProtected screenssign-in, payment,“Delete for good”,sharing, keys…no design reaches themA second wall: the app only loads pictures and fonts from itself,so even a rule that slipped past the checker cannot send anything out.
Looks are safe by construction. Code lives in sealed boxes with only the doors you open.
  • Settings, CSS and HTML are read into the app’s own form and checked against a list of what is allowed. They cannot run code and contain no web addresses. See HTML and CSS rules.
  • Checked twice. The engine checks a design when it is saved, whoever wrote it: you, the AI, kzx, an MCP client such as Claude Code, or an editor on your folder. A design that breaks a rule is not saved, and the message names the line. The screen checks it again before it paints. Example: a skin whose CSS loads url(https://…) is refused with “url() may only name an uploaded file”.
  • Scripts run in a sealed engine: a separate, locked JavaScript engine with no screen, no internet, no files and no keys. A script shows what it reads, writes and calls before it runs, and cannot do more.
  • HTML blocks run in a sealed frame that cannot see the app, your keys or your storage. Said plainly: a frame can still send out what it sees through WebRTC (a video-call feature no browser lets a page block). So before you turn on an HTML block someone else made, the app shows exactly what it can see. Your hidden collections are never offered to it.
  • Uploaded files (images, GIFs, SVGs, fonts, Lottie animations) are checked and cleaned, and served from a separate address. A design never uses a web link for a picture.

Protected screens

These screens cannot be styled at all: sign-in, sign-up, password reset, onboarding, payment and plans, every permission and “Enable” screen for code, Safe mode, Reset, keys and recovery, end-to-end encryption choices, removing a device, the AI key, disconnecting a connector, unpublishing, “You’re leaving Kaizendex”, the share dialog, every “Delete for good”, deleting your account, the mass-delete pause card, the AI’s approval buttons and receipt, and the settings pages for security, plan, devices, privacy, where your data lives, AI routes and AI safety.

They have a fixed look and the system font. No design can dim, move, cover, hide, restyle or relabel them. Not even Free CSS.

How they are walled off: each protected screen is drawn inside a sealed part of the page (a closed shadow root). Outside stylesheets cannot enter it, so no design rule reaches it, whatever it aims at. Inside, the screen uses the app’s built-in theme, light or dark as you chose, and the system font; your own theme does not apply there either.

The screens you act on (a protected dialog such as “Delete for good”, and the AI’s approval card) go one step further. They are drawn in the browser’s top layer, a level above every other part of the app, outside the page itself. No design can hide them, cover them or lay a box over them. A design also cannot use position: fixed, or lift a ::before or ::after out of the flow (absolute, sticky) to float a box over the app, and a Look cannot set pointer-events (a box that lets clicks through can still cover what is under it). Only a skin may, because a skin cannot leave its own part.

The protected Settings pages stay in their place on the page. If a design hides or covers one of them, the app turns every design off at once (Safe mode) and tells you why. The app checks this by looking at the page with the designs and again without them, so an open menu of the app’s own never counts. It checks when the page appears, and again on every resize, scroll and animation, when the section scrolls into view, and every second while it is open.

Example: a Look sets every dialog to opacity: 0. Your normal dialogs fade away, but “Delete for good” still shows, with its usual look.

Today this covers “Delete for good”, the AI’s approval card, Safe mode and Find the problem, Reset all designs, and the settings pages for plan, AI key and routes, security, where your data lives, devices and deleting your account. Menus, tooltips and dialogs opened from a protected screen are walled off from designs too.

Other parts are partly locked: a dialog’s buttons keep their order, wording and font; a toast is always shown and readable; a menu item that deletes keeps its place and wording.

Who made it decides what is on

The danger in someone else’s design is never its looks. It is what it can do.

Made by Looks (settings, CSS, HTML) Doing things (a button’s change, links, AI questions, scripts, HTML blocks)
You, in the app on on
Your AI, in the app on, as a draft you keep or undo wait for your Approve
The Library (you copied it) on after the checks off until you turn each on
Another member of a shared space wait for your Use it off
A file in your folder changed by another program wait for your Use it off until approved; any change asks again

Why Library looks are on but shared-space looks wait: a fake “type your password” box only works if what you type gets back to the person who made it. A Library design has no way back. A shared space does: what you type into a shared entry syncs to the other members.

Approving a script or an automation is a change only you can make: you approve it in the app, on your device. The AI, your folder and other programs can only suggest it. An HTML block someone else made starts off too. You switch it on in a protected Enable screen that shows exactly what the block can see.

  • An HTML block that runs a script follows the same rule as a script. Approving it is a change only you can make: the app asks for Face ID, a fingerprint, Windows Hello or your device password first.
  • An HTML block with no script is only HTML and CSS in its sealed frame. Approving it is not a protected change: your tap on the Enable screen is enough.

The AI never writes code inside an automation that runs while you are away. It may have read a hostile web page or message.

Actions that change your data (set, addTag, timer in a Look’s on) are stricter still. They are off in every design, yours too, until you allow them for that design in Settings → Designs → Actions that change entries. That list is a protected screen, so no design can restyle or cover it. Only you add or change them. While a design has one, the AI, kzx, MCP, a file edited in your folder, and Undo by any of them can change only how it looks (style, states, CSS, tokens, background). Anything else they try is refused.

Your designs stay with you

  • Your designs are stored on your own devices and in your own Cloud Sync copy. They run only on your devices. Only automation scripts can also run in the cloud: on Pro, in your own part of Cloud Sync, while your devices are off.
  • Nothing you write runs on Kaizendex’s shared servers. Library previews are drawn on a separate web address.
  • So a mistake only affects you, and Safe mode turns every design off in one click.

Free CSS

Free CSS is an opt-in for your own account. Your CSS may then aim at any part, even one without a name tag. Every other wall stays: no web addresses, locked dialog buttons, and protected screens out of reach. It is never shared through the Library.