MCP tools
MCP (Model Context Protocol) is a standard way for an AI app to call tools in another app.
Kaizendex’s MCP server is part of kzx. It lets agents without a terminal, such as Claude
desktop, use your data with the same rules as the app.
Where it runs
| Where | Status |
|---|---|
Your computer, through kzx (with or without the desktop app) |
first release |
| A personal MCP address for cloud agents, through Cloud Sync, limited to what you allow | later, Pro |
| Web app and phones | no |
Does it work in my setup?
Example: Claude desktop calls search for “knee pain” in your Workouts.
| Your setup or your data | kzx and MCP on your computer |
A cloud MCP address (later) |
|---|---|---|
| Free, Cloud Sync on | yes | no: it needs Pro |
| Pro, Cloud Sync on | yes | yes |
| Cloud Sync off (desktop app only) | yes | no: it needs Cloud Sync |
| A Cloud Sync collection | yes | yes, on Pro |
| A Cloud Sync, end-to-end encrypted collection | yes | only with a key you give that connection |
| A This device only collection | yes, on the device that holds it | never |
| A No AI collection | refused | refused |
| Hidden entries | refused | refused |
- On your computer,
kzxworks the same in every setup: Cloud Sync on or off, end-to-end encryption on or off. Your own computer can already read your own data. End-to-end encrypted means only your devices hold the key, and your computer is one of them. - Your privacy choices still apply. “No AI” collections are never given to AI tools. “Ask me first” waits for your Approve. Locked files stay locked. Hidden stays hidden.
- A cloud MCP address (later) lets cloud AI tools, such as Claude Code on the web, reach your data while your computer is off. It reads your data in Cloud Sync, so it needs Pro.
Tools
| Tool | What it does | Example |
|---|---|---|
search |
words and meaning together | “knee pain” |
search_entries |
structured search: collection, property values, dates | Workouts over 10 km in October |
recent |
the newest entries | the last 5 Sleep entries |
related |
entries close to one entry | notes like “Sync plan” |
read |
one entry, in full | “Morning run” |
context |
an entry and its whole neighbourhood, to a chosen depth and size budget (as kzx context) |
a project with its tasks and notes |
list_skills |
lists your skills | — |
read_skill |
reads one skill | weekly-review |
list_automations |
lists your automations, on or off | — |
set_automation |
makes or changes one automation (needs the AI permission below) | “when Done is checked, set Stage to Done” |
set_automation_enabled |
switches one automation on or off (needs the AI permission) | switch on “Done pile” |
delete_automation |
deletes one automation; Undo brings it back (needs the AI permission) | delete “Done pile” |
delete_collection |
moves a whole collection and its entries to the Trash, as one change; restore_item or Undo brings it back (needs the AI permission “AI can move collections to the Trash”) |
move “Old reading list” to the Trash |
These search tools have no AI inside them. They are plain and fast; the calling agent decides what to call next.
Rules
| Rule | What happens |
|---|---|
| Every caller | treated as an AI, even when you call a tool yourself |
| A collection set to No AI | refused |
| A collection set to Ask me first | the app shows an Approve card that names the program (“An MCP client calling itself “claude-desktop” wants to read Diary”; the client picks its own name); the call waits |
| A change only you can make: Where, AI choices, sharing, keys, approving an automation, AI safety | the call only suggests it; the suggestion waits in the app until you approve it on your device |
A write to an automation (set_automation, set_automation_enabled, delete_automation) |
refused unless you ticked “AI can make and switch on automations” in Settings → Security → AI permissions; the refusal names the permission and carries its link (permission.link: kaizendex://settings/security/ai-automations), which opens the checkbox in the app, highlighted. See Safety and approval |
delete_collection |
refused unless you ticked “AI can move collections to the Trash” in Settings → Security → AI permissions; the refusal carries its link (kaizendex://settings/security/ai-collections). A collection closed to AI, or one holding entries the AI cannot read, is always refused. Restoring needs no permission |
| Hidden entries | refused |
| End-to-end encrypted data, over a personal MCP address (later) | refused unless you give that connection a key for the collection |
| This device only data, over a personal MCP address (later) | never: it is not in the cloud |
| Every write | checked like a tap, logged with the agent’s name, undoable |
The AI in the app, kzx and this MCP server run the same tools, with the same rules. A new tool
reaches all three at once.
Outside MCP tools your agents use
This page is about tools Kaizendex offers to other apps. It works the other way too: your Kaizendex agents can call tools from another app’s MCP server (later, in the agent home).
| Rule | What happens |
|---|---|
| A call to an outside tool | goes through the outbox, unless you marked the tool read only |
| A tool that changed since you approved it | blocked until you approve it again (by default) |
| A cloud MCP server | runs in its own sealed container, apart from your data |
| A tool marked read only | skips the outbox, but can still leak what it reads: its maker gets what the AI sends it |
Example: you approved a “Strava” tool for Coach. Its maker adds a new “post activity” action. By default, the tool is blocked until you look at the change and approve it again. The rules marked “by default” are your AI safety settings.