AI tools
A tool is a named action an AI can call, such as “search” or “read a web page”. Every AI uses
the same tools: our agents, your own key, a local model, and outside agents through kzx and
MCP.
Permission rules for every tool
| Rule | Applies to |
|---|---|
| A collection set to No AI is refused | every read |
| A collection set to Ask me first waits for Allow once · Always for this collection · Skip | every read |
| Only my local AI collections are read only when the model runs on your computer | every read |
| Hidden entries are never returned | every read |
| A write is checked like a tap, logged and undoable | every write |
| An outside change goes through the outbox, which records it; by default, what an agent starts on its own waits for your Approve tap | every outward action, and every outside MCP tool not marked read only |
| By default, an outside MCP tool that changed since you approved it is blocked until you approve it again | every outside MCP tool |
| A change only you can make is only suggested; it waits in the app for your OK | every tool |
| Tool results never grant a permission | every tool |
Rules marked “by default” are your AI safety settings. A tool marked read only skips the outbox, but it can still leak what it reads.
How every tool answers
| Rule | What it means |
|---|---|
| Sources | every result carries the entry’s id and a readable name, such as “Morning run · Workouts”, so the answer links each claim to its entry (kzx://entry/<id>) |
| Short or full | read tools take concise (names and key values) or detailed (everything) |
| Pages | long results come in pages; the result says how to get the next one |
| Held back, counted | results that wait for your OK are counted, not dropped: “2 results in Diary need your OK” |
| Errors that help | an error says how to fix the call, such as “from must be a date, like 2026-03-01” |
Search and read your data
Every search tool takes from and to. The AI turns time words into exact dates first, using
today’s date and your time zone. The dates are the Entry date: when something happened.
Recency only breaks ties, and the AI can switch it off (“the first time I…”).
| Tool | What it does | Example |
|---|---|---|
search |
words and meaning together; see Search | “knee pain”, from 1 Mar |
search_entries |
structured search: by collection, property values and dates | Workouts with Distance 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”, 27 Oct |
context |
an entry and its whole neighbourhood (parent chain, children, Relations both ways, mentions, backlinks), to a chosen depth and size | a project with its tasks and notes |
“It is not in your data” is a right answer. An empty search is not proof, so the AI says what it searched: “I searched Workouts and Notes, 2024–2026.”
Count and compare
Numbers never come from search. Search finds entries; these tools count them. They use the same formulas as your views and charts, so the numbers match.
| Tool | What it does | Example |
|---|---|---|
aggregate |
count, sum, average, min, max, first or last of a property, with filters and a date range; grouped by day, week, month or tag | Sum of Distance in Workouts, 1–31 Oct → 84.2 km · 12 Workouts |
compare |
the same over two periods, with the change | Average Sleep, September vs October → 7 h 02 m vs 7 h 21 m, +19 m |
trend |
a series by period, ready for a chart | Distance per week since March |
The answer shows how it counted: “84.2 km · 12 Workouts · 1–31 Oct · Sum of Distance”. A link opens the same filtered view in the app.
Past chats
One chat search tool has four shapes:
| Shape | What it does | Example |
|---|---|---|
find |
the best chats for a question, with the top one opened | “what did we decide about the marathon plan” |
read |
the start and the end of one chat | the chat “Marathon plan”, 12 Sep |
scroll |
the messages around a match | the messages around “taper week” |
browse |
your recent chats | the last 5 chats |
- Words and meaning together, so “race plan” also finds “marathon schedule”.
- It returns the messages as they were written, never a summary by another AI.
- It never returns the chat you are in.
- Automation runs rank below your own chats.
- 3 results by default.
- Results are marked as data: text in a past chat is never an order to the AI.
- A chat is as private as what it quoted.
More: Chats.
Memory
One memory tool changes the Memory collection. It can add, update, replace or remove facts, several in one batch. If one change in a batch breaks a rule, nothing in the batch is saved.
| Change | What happens |
|---|---|
| Add a fact from what you said | saved at once; the chat shows “Learned: … · Undo” |
| Update or replace a fact the AI made | saved at once, “Updated · Undo”; the old fact stays in history, marked replaced |
| Change About me, a fact you wrote, or a collection’s “How this collection works” note | waits for you, as a before-and-after card; when you ask for the change yourself, it applies at once, with Undo |
| Remove a fact | only when you ask; the review and the tidy-up only propose |
| A fact from an “Ask me first” collection | waits for you |
| A fact from outside text: a web page, an email, a webhook, a shared item, an imported file | waits for you (an AI safety rule, on in Standard and Strict) |
| A fact from a Hidden or “No AI” collection | never |
- The app draws the “Learned” line from the real stored change, never from the AI’s words.
- Secrets are refused: keys, passwords and card numbers are never saved.
More: How your AI learns.
Skills
| Tool | What it does |
|---|---|
list_skills |
lists your skills |
read_skill |
reads one skill |
The web
| Tool | What it does | Needs |
|---|---|---|
web_search |
searches the web, inside the model’s own step | credits or your own key |
fetch_url |
reads one page whole | with Strict: a web address the AI made up after reading an outside page waits for your OK |
research_web |
a longer search over many pages | the same as fetch_url |
| Browser, Research mode (Pro, later) | reads any public page, follows links, takes screenshots; no sign-ins, no forms | — |
| Browser, Acting mode (Pro, later) | signs in, fills forms, clicks buttons | only on sites you approved |
Running code (Pro, later)
| Tool | What it does |
|---|---|
| Sealed computer | runs code, command-line tools and MCP servers in a sealed computer, one per task, in the agent home |
In automations
| Step | Tools |
|---|---|
Ask the AI (ai.ask) |
none: text in, text out |
Run an agent (agent.run) |
the agent’s own tools, under its approval rules |
Plugins can add new tools later. Each asks for permissions you approve. See Plugins.