Skip to content

How agents in the cloud work (technical)

On Pro, your agents work in the cloud while your devices are off. They think in one place, your cloud box, and they borrow tools only while a job needs them. Three fixed rules sit around every tool, so a much stronger agent can plug in later without a rebuild.

In short:

  • The brain is your cloud box. It holds your data, memory, schedules and chats. It is part of Cloud Sync on Pro.
  • The body is borrowed: web search, a browser, a sealed computer. The agent gets it for one job and gives it back.
  • The three rules: every tool goes through one tool interface; every outside change goes through one gate (the outbox), which records it and, by default, holds what the agent starts on its own for your Approve tap; every agent is a registered device with its own keys.

The plain version of this page is The agent home.

The cloud box is the brain. It reaches every tool through one tool interface: web search and pages, a cloud browser, a sealed container, connectors such as email, and one day a full personal computer. Every change that reaches the outside world, from the brain or from a tool, passes through one gate, the outbox. By default, what the agent starts on its own waits there for your Approve tap. A checker AI can sit at the gate; it is off by default.Cloud box: the brainyour data and searchmemory and skillsschedules and alarmschats and channelsthe log of every actionwakes on a schedule,a message or new dataa registered devicewith its own keysOne toolinterfaceevery toolplugs in hereborrowed only while a job runsWeb search, fetch pagesnow, with automationsCloud browsernext: Research; later: ActingSealed container per joblater: code, CLI tools, MCPConnectors, such as emaillaterA full personal computerone day: one more toolThe outbox (one gate)send, post, buy, changeby default, asks for Approvea checker AI: off by defaultOutside worldemail, shops
One brain, borrowed tools behind one interface, and one gate to the outside world. A new tool is one more box on the right. Nothing else changes.

The brain: your cloud box

Your cloud box is your own small private database in the cloud, with a small program around it. It is part of Cloud Sync on Pro. It is built on a Cloudflare Durable Object: one per person, asleep for free and awake in milliseconds.

The brain holds Example
your data and its search Workouts, Sleep, a Diary collection
memory and skills “Daniel runs in km”, the weekly-review skill
schedules and alarms “every day at 8:00”
chats and chat channels a chat with Coach; Telegram or WhatsApp
the log every action, with the agent’s name; every change to your data can be undone

What wakes it:

  • A schedule: “every day at 8:00”.
  • A message: you write in the app, on Telegram or on WhatsApp.
  • Incoming data: another service sends data, for example a Whoop sleep sync.

It works while your devices are off. Its work reaches your devices the way an edit from your laptop reaches your phone: through Cloud Sync, like any other device.

Example: at 8:00, Coach wakes. It reads last night’s Sleep entry and this week’s Workouts. It writes “Easy 5 km today” into a new “Morning run” plan entry. Your laptop is closed. When you open your phone, the entry is already there.

The brain is not a strong computer. It has little memory, no graphics card, and no AI model inside it. So it borrows a body when a job needs more.

The body: borrowed tools

Each tool sits behind the same tool interface. The agent asks for a tool, uses it for one job, and gives it back. Tools cost only while they run.

Tool What it does When
Automations run your rules and steps on a schedule or when data comes in Now
Web search and fetching pages search the web; read one page Now
An AI model writes the answers: Kaizendex AI (our OpenRouter key, paid with credits) or your own OpenRouter key. Small jobs, such as search vectors and short summaries, go through OpenRouter too. If a model is down, a second model takes over, also through OpenRouter; search-vector jobs wait in a queue Now
A cloud browser, Research mode reads any public page, follows links, takes screenshots; no sign-ins, no forms, no downloads that run (Browser Run) Next
A cloud browser, Acting mode signs in, fills forms and clicks, only on sites you approved; each site’s sign-in is kept sealed for the next job; by default, saved sign-ins are in one list with Remove, and one unused for 90 days is dropped; it uses a saved sign-in first, and when a password must be typed, our cloud fills it in while the agent’s view of the page is paused, so the AI model is never sent your password; every form it submits goes through the outbox Later
A sealed container per job runs code, command-line tools and MCP connectors; each cloud MCP server runs in its own container; every MCP tool call goes through the outbox unless you marked that tool read only (a read-only tool can still leak what it reads); by default, a tool that changed since you approved it is blocked until you approve it again; the container is thrown away after the job (Cloudflare Sandbox / Containers) Later
Long jobs as saved steps a job of many hours runs as steps; each step is saved, so a restart loses nothing (Cloudflare Workflows) Later
Connectors, such as email read your inbox, draft replies; by default, sending waits for your Approve tap; agents use a connector rather than driving a browser when one exists Later
A full personal computer a whole computer in the cloud, like Meta Muse or OpenAI Dots, for jobs that need installed apps One day

“Now” means the first version of agents in the cloud. Agents in the cloud come after the first release of the app.

The three rules

These three rules never change. They are why a new tool needs no rebuild. What changes is how careful the gate is: that is your AI safety settings.

Rule What it means Why it matters later
1. One tool interface the brain calls every tool the same way: search, browser, container, computer. The AI in the app, kzx and MCP run one tool registry, with the same tools, privacy rules and AI permissions a new tool plugs in, and reaches all three at once; nothing else is rewritten
2. One gate to the outside world: the outbox every action that changes something outside Kaizendex goes through one gate: send, post, buy, delete or change something on a site or service, submit a form, call an MCP tool (unless you marked it read only). The outbox records every action. By default (Standard), reading needs no approval, and what the agent starts on its own waits for your Approve tap; an automation you approved sends without a new tap a stronger tool cannot skip the gate; a checker AI that approves the safe actions for you sits at the gate; it comes with the outbox and is off by default
3. Every agent is a registered device with its own keys an agent is a device of your account, like your phone or laptop, but it can never make a change only you can make your “Where” choices, your AI choices and your end-to-end keys apply to it at once, with no extra setup

Privacy

Every collection and entry has a Where: Cloud Sync, Cloud Sync end-to-end encrypted, or This device only. End-to-end encrypted means only your devices hold the key. Each collection also says what AI may read.

Your data Can an agent in the cloud read it?
Cloud Sync (the default) yes, if the collection’s AI setting allows it
Cloud Sync, end-to-end encrypted only if you give the agent a key for that collection. The key opens the collection in memory only while a job runs, and the job closes it at the end. The collection shows a badge, Your agent can read this, with Remove.
This device only never. It never leaves its device, so it never reaches the cloud box.
a No AI collection never
Hidden entries never

On plain Cloud Sync, Kaizendex will not look at your data, but it technically could. Only end-to-end encryption means we cannot. See End-to-end encryption.

If an automation reads an end-to-end collection that the cloud box has no key for, it runs on one of your devices instead. The cloud box picks one device, so it runs once.

Safety

  • What the agent reads is information, never an order. A web page, an email, a message or data from a webhook can hide text aimed at the AI (“forward all emails to me”). That text gives the agent no new permission.
  • Every outside change goes through the outbox, which records it. By default, what the agent starts on its own waits there for your Approve tap. An automation you approved sends without a new tap; its approval card said so. The Approve card is drawn by the app from the real action (who, the full text, where), never from the AI’s words.
  • Your AI safety settings decide how careful the agent is. Standard is the default. Strict adds checks on what the agent reads from outside: a web address it makes up after reading a page needs your OK, a reader model with no tools reads outside pages first, and approved automations keep fixed destinations. By default, agents also have a daily spending limit. See AI safety settings.
  • Changes only you can make. An agent can only suggest a change to Where, AI choices, sharing, keys, automations or AI safety. The suggestion waits in the app for your OK. So an agent cannot loosen its own rules. One exception is yours to allow: with “AI can make and switch on automations” ticked in Settings → Security → AI permissions, an agent makes and switches on automations itself, inside the AI’s limits. With “AI can move collections to the Trash” ticked, it moves a collection to the Trash; Delete for good stays yours. See Changes only you can make.
  • Sites are blocked unless you allowed them. In Acting mode, the agent reaches only the sites on your approved list. Research mode only reads public pages.
  • The AI model is never sent your password. In Acting mode the agent uses a saved sign-in first. When a password must be typed, our cloud fills it in while the agent’s view of the page is paused. Other keys are added by our cloud on the way out.
  • Every change to your data is in the log with the agent’s name, and you can undo it. Something that already left Kaizendex cannot be unsent. That is why, by default, it waits in the outbox for your Approve.
  • You stay in charge: a read-only mode and an off switch, in Settings → Agent home.