Skip to content

AI safety settings

AI safety settings decide what the AI and your agents must ask you before they act. You find them in Settings → AI → Safety. Pick a preset, or set each rule yourself.

Coach wants to email your physio. Your Safety rules (Standard) say ask first, so the app shows an Approve card drawn from the real action. After your tap, the email is sent and recorded in the outbox. Only you can change the Safety rules, in the app.Coachwants to emailyour physioSafety rulesStandard:ask firstApprove carddrawn by the appto · text · whereSentrecorded inthe outboxOnly you change the Safety rulesin the app, on your device. An AI can only suggest a change, so it cannot loosen its own rules.
Your Safety rules decide when the AI asks. The outbox records every outside action either way.

The presets

Preset What it does Good for
Standard (the default) the AI asks before it sends, posts, pays or deletes outside the app, and the main guards are on chats and agents that work mostly on your own data
Strict Standard, plus three more checks on what the AI reads from outside agents that read many web pages, emails or messages
Custom you tick or untick each rule yourself, every rule included a setup of your own

Each rule has a checkbox. Ticked means on.

The rules

Rule What it means Standard Strict
Ask before acting outside the app the AI asks before it sends, posts, pays or deletes anything outside Kaizendex; the action waits in the outbox for your Approve on on
Block changed outside tools an outside tool (an MCP tool: a tool from another app) that changed since you approved it is blocked until you approve it again on on
No link previews in chat channels agent replies on Telegram or WhatsApp show no link previews and load no images by themselves on on
Skills never grant powers a skill from the Library gives the AI no new tools or permissions; an update shows you its changes first on on
Daily spending limit for agents agents stop for the day when they reach a spending limit on on
Saved sign-ins in one list every site your agent stays signed in to is in one list, each with Remove; a sign-in unused for 90 days is dropped on on
Outside text cannot teach the AI anything the AI would learn from a web page, an email, a shared item or an imported file waits for your OK before it becomes memory or a skill. See How your AI learns on on
Ask before new web addresses after the AI reads an outside page, a web address it makes up itself waits for your OK; a link you gave it, or one it found in search or on a page, does not off on
Reader model for outside pages a separate AI with no tools reads each outside page first and passes on only its content, as plain data off on
Fixed destinations where an approved automation sends things (an email address, a website) is fixed when you approve it; what it sends may change each run off on
Checker AI a second AI approves the safe actions for you; you still approve the rest off off

Example: Coach reads a race page, then wants to open a web address it wrote itself. On Standard it goes ahead. On Strict it waits for your OK first. A web address can carry data out of the app, which is why Strict asks.

The checker AI comes with the outbox. It is off in both presets. You can tick it in Custom.

Always true, whatever you pick

  • Only you can change Safety. It is a change only you can make: in the app, on one of your devices. An AI, an agent, kzx or MCP can only suggest a change, and it waits for you. So an AI cannot loosen its own rules, even when a web page tells it to.
  • The outbox records every outside action. Every send, post, payment or delete is listed, with Undo where possible, even when a rule does not ask you. A sent email cannot be unsent.
  • The Approve card shows the real action. The app draws it from the action itself: who it goes to, the full text, and where. It never uses the AI’s own words about the action.
  • Agents prefer a connector. When a service has a connector, such as email, an agent uses it instead of driving a browser.
  • Outside tools in the cloud run apart. Each cloud MCP server runs in its own container (a sealed computer of its own).

AI permissions

Safety decides what the AI must ask before it acts. AI permissions decide what an AI may do at all. You find them in Settings → Security → AI permissions: one checkbox per power. Ticked means allowed.

Checkbox Off (the default) On
AI can make and switch on automations only you make, change, switch on or delete an automation, in More → Automations; the AI can suggest one an AI makes, changes, switches on and off, and deletes automations; they run with no Approve card
AI can move collections to the Trash only you delete a collection; the AI can suggest it an AI moves a whole collection to the Trash, with its entries; Undo or the Trash brings it back

“An AI” here means an agent in a chat, kzx, and MCP apps such as Claude Code and Codex.

Example: you tick the checkbox and ask your AI: “When Done is checked on a task, set Stage to Done.” It makes the automation, and it runs.

Another: you tick AI can move collections to the Trash and ask Claude Code to remove your empty collections. It moves each one to the Trash. Changed your mind? Undo, or restore it from the Trash.

  • Only you tick a permission, in the app. It is a change only you can make. An AI can never read or change it, so it cannot give itself a power.
  • When a permission stops an AI, the answer names the permission and links to it. The AI tells you: “I can do that if you allow it.” In a chat, a card “Kaizendex needs your permission” has one button, Open AI permissions. The button (or the link from kzx, kaizendex://settings/security/ai-automations, or …/ai-collections for collections) opens Settings → Security with the checkbox highlighted. You tick it there; nothing is ticked for you.
  • An automation an AI made keeps the AI’s limits until you switch it on yourself. It never reads a collection closed to AI and never runs an agent. See Safety and approval.
  • Safety still applies. An automation an AI made sends outside the app only as your Safety preset allows.
  • Some collections are never moved by an AI: one closed to AI (No AI, Hidden), or one that holds entries the AI cannot read. Restoring a collection from the Trash needs no permission. Delete for good is always yours.
  • The same everywhere. The AI in the app, kzx and MCP use one set of tools, with the same privacy rules and the same permissions. A tick here counts for all three.

More permissions will join this list later, one checkbox each. A safety rule is a default you can change, never a wall with no door.

The rules are ordinary rules

The Safety rules are ordinary rules in the app’s one JSON language, the same language as your views, card designs and automations. You can read them, change them, copy them and share them like any block. Example: a friend shares a Custom setup with Strict plus the checker AI. You copy it and read every rule before you use it.

Using copied rules as your own Safety is still a change only you can make. With the desktop app, your settings are also files in ~/Kaizendex/Settings/. An edit to the Safety rules there from another program waits for your OK in the app.