Safety and approval
You approve an automation once; after that it runs by itself, and it can only do what you approved.
Approve once, then it runs
Before an automation can run, the app works out its inventory: the list of what it reads, what it changes, which websites it calls, whether it asks the AI or runs an agent, and what it sends. The app reads this from the code itself. It never takes it from the automation file, so nobody can write a false list.
- You approve on the protected Enable screen. A protected screen is one no design can restyle, cover or fake.
- When the AI makes an automation for you, there is no approval card. The AI makes it only if you allow it, in Settings → Security → AI permissions (see below). Then it runs.
- After that it runs every time by itself, with no more questions. That includes what it sends: a message or a website call in an approved automation goes out without a new tap. The Enable screen says so.
- It asks again only when an edit makes it read, change, call or send more, or when a run reads a collection you set to “Ask me first”.
Approving an automation, and the powers it gets, is a change only you can make. You approve in the app, on one of your devices, after a quick check (Face ID, fingerprint, Windows Hello or the device password). Your device signs the approval. Your other devices and the cloud check that signature before they run the automation. Your folder and other automations can never approve. An AI can switch an automation on only when you allow it, and then the automation keeps the AI’s limits (see below).
Let the AI make automations
You decide if an AI may make automations. The checkbox is in Settings → Security → AI permissions: “AI can make and switch on automations”. It is off by default.
| The checkbox | What an AI can do |
|---|---|
| Off (the default) | It reads your automations, to explain one or to suggest a new one. It cannot make, change, switch on, switch off or delete one. |
| On | It makes, changes, switches on, switches off and deletes automations. They run at once, with no Approve card. |
“An AI” means an agent in a chat, kzx, and MCP apps such as Claude Code and Codex.
Example: you tick the checkbox. You ask Claude Code: “When Done is checked on a task, set Stage to Done.” It makes the automation and switches it on. The next task you tick moves to Done. History shows one change, written as “Claude Code”. One Undo takes it back.
When the checkbox is off, the app says no. It tells the AI which permission is missing, and gives a link to it:
“Automations/Done pile” is an automation, and AI is not allowed to make, change, switch on or delete automations; mcp cannot. The person can allow it in the app: Settings → Security → AI permissions → “AI can make and switch on automations” (open it: kaizendex://settings/security/ai-automations). Until then, suggest it: tell the person what it should do, and they make it in More → Automations.
The AI passes this on to you: “I can do that if you allow it.” Then you pick: tick the checkbox, or make the automation yourself in More → Automations.
- In a chat in the app, the agent says so in words. Under its message, a small card says “Kaizendex needs your permission”, with one button: Open AI permissions.
- From
kzx, the answer ends withOpen the permission: kaizendex://settings/security/ai-automations. An agent such as Claude Code shows you that link. - The button and the link open the desktop app on Settings → Security. The app scrolls to the checkbox and highlights it for a moment. You tick it there. Nothing is ticked for you.
Every AI permission works this way: when one stops an AI, the answer names it and links to it.
What stays safe when it is on
- Only you tick the checkbox. It is a change only you can make. An AI can never read or change it, so it cannot give itself the power.
- An automation an AI made keeps the AI’s limits (below), until you switch it on yourself in More → Automations. Then it is yours.
- Your folder and other automations never write automations.
Settings/Automations/stays read-only. An edit there is not applied. It is listed inproblems.md, and the file keeps its text. - Every change is one step in History, with the name of the AI that made it. One Undo takes it back.
- Sending outside the app still follows your AI safety settings. In Standard, a message or website call waits in the outbox for your Approve.
Your automations from today’s app come over with the import. They count as yours.
What a rule may touch
An automation you made runs as yours. It may read and change the collections it works on, the private ones too, under the privacy rules below.
An automation that no person made runs with the AI’s limits. That is an automation an AI made with your permission, or one an older build let a program write. The app reads who wrote each version of a rule from its history. Such a rule:
- never reads or changes a collection closed to AI (No AI, Hidden or Ask me first), and never learns the titles of entries there;
- never runs an agent;
- fails with one plain line in its run history: “… was not made by you in the app …”.
Example: a program once wrote “Diary page”, a rule that makes a page in your Diary each night. Diary is set to No AI, so the rule does not run, and its history says why. You open it in More → Automations and switch it on yourself. Now it is yours, and it runs.
Agent steps
A step that runs an agent must name one of your agents: an entry in your Agents collection. Its model must be empty (then the agent uses your chat model) or a model you chose in Settings → AI. If not, the run fails with one plain line, and no agent run is queued.
Example: “Morning motivation” runs The Navy Seal with no model set, so it uses your chat model. A rule that names an agent called “ghost”, or a model you never chose, does not run.
Who made it decides if it starts on
| Made by | Everyday actions | Scripts |
|---|---|---|
| You, in the app | on | on, after you approve its inventory |
| Your AI (only with AI can make and switch on automations ticked) | on when the AI switches it on; runs with the AI’s limits until you switch it on yourself | the same |
| The Library (you copied it) | off until you switch it on | off until you approve it |
| A shared space | off until the owner switches it on for the space, and you opt in to it changing your entries | off |
| A file in your folder changed by another program | off until you approve it in the app | off until you approve it |
Today an edit in the folder is not applied at all. The old version stays on, and the edit is
listed in problems.md.
If a Library automation turns out to be bad, we switch it off on every device that copied it.
Privacy decides the whole run
A run is as private as the most private thing it reads. Before a run starts, the app checks every collection in the inventory.
| If the run reads anything set to… | Then, for that run |
|---|---|
| Any AI only | everything is allowed |
| Ask me first | the run waits for you: “Weekly review wants to read Health · Allow once · Always · No” |
| Only my local AI | AI questions go only to your local model; websites, messages and cloud AI are off |
| No AI, Hidden or locked | no AI, no websites, no messages; a notification says only “Weekly review finished”, never content |
What a run writes follows the same rule. A run that read private data can only write into collections that are at least as private, or the write waits for your OK. So an automation cannot copy your Diary into a collection the AI can read. Its run history and any summary it writes are stored as privately as the most private thing it read. More on the choices: What AI may read.
Websites and secrets
- An automation lists the exact websites it may call (
httpsonly, one exact address, no wildcards). You approve the list. A redirect to anywhere else is refused. - Private and local addresses (your router,
localhost, your local AI server) are always refused. - The app makes the call for the script and hands back only the answer. The answer is data, never an order to the AI.
- A secret (for example a key for a weather service) is tied to one website. It is kept in your keychain, or on Pro in a vault inside Cloud Sync. The app adds it to the request. The script never sees it.
- For an automation from someone else, the Enable screen says in plain words what leaves: “sends data from Workouts to api.example.com”.
- Fixed destinations is a rule in your AI safety settings: on in Strict, off in Standard. When it is on, where an approved automation sends things (an email address, a website) is fixed when you approve it. What it sends may change each run. Where it goes may not. Example: “Weekly review” emails your physio every Sunday. With Fixed destinations on, it cannot email anyone else until you approve it again.
Your AI safety settings
What the AI and your agents must ask before they act is a setting, in Settings → AI → Safety. Standard is the default. Strict adds more checks, such as fixed destinations. Custom lets you tick or untick each rule. Whatever you pick, the outbox records every outside action, and changing Safety is a change only you can make. See AI safety settings.
Two walls around every script
- A sealed engine. Each run gets a fresh, small JavaScript engine (QuickJS) inside WebAssembly, a closed box. It has no internet, no files, no keys, no cookies and no app storage. It has only the toolbox. A read outside its inventory returns nothing.
- Your own devices and your own part of the cloud. Scripts run only on your devices and, on Pro, in your own part of Cloud Sync, which holds only your data and keys made for you alone. They never run on Kaizendex’s shared servers.
ai.ask in a script has no tools, so text inside your entries cannot make the AI open a page or
act. An automation never deletes and only adds text to the end of a Block page property.