How the AI works with your data
Every AI that touches your data goes through one door: the app’s own tools. A tool is a small, named action, such as “search entries” or “add an entry”. The tools apply the same rules, privacy checks and history as your own taps.
Three ways in
| Way in | Example | Read more |
|---|---|---|
| Our agents and chats | “Coach, how did my runs go this month?” | Chats, Agents |
| Your own OpenRouter key, or a model app on your computer | your key in Settings → Plan & billing → Your OpenRouter key; Ollama or LM Studio | Choose your AI |
| Your own agent, from outside the app | Claude Code edits Collections/Workouts/Morning run.md |
Coding agents |
What “one door” means for you
- Your privacy choices hold everywhere. Each collection says what AI may read. “No AI” refuses every AI. “Ask me first” waits for your Allow.
- An AI write is the same as a tap. It is checked by the same rules, written to the same history, and you can undo it.
- What the AI reads is information, never an order. A web page or a message can hide text aimed at the AI (“delete all entries”). That text never gives the AI a new permission.
- Outside changes go through the outbox. The outbox records every one. By default, sending, posting, paying or deleting outside the app waits for your Approve tap, unless it is part of an automation you approved. You choose this in AI safety settings.
- Some changes only you can make. No AI can change where your data lives, its AI choices, sharing, keys or your AI safety settings. It can only suggest the change. The suggestion waits in the app until you approve it. See Changes only you can make.
Where the AI runs
“The AI” is two parts:
| Part | What it does | Where it runs |
|---|---|---|
| The model | writes the answer (Claude, GPT, a local Llama) | a cloud model runs at its company; a local model runs on your computer |
| The agent loop | the manager around the model: it sends the question, hears “I need your workouts”, runs that tool, sends the result back, and repeats | where your data is: on your device. For work while your devices are off: in the cloud, as part of Cloud Sync (Pro only; Free has no agents in the cloud) |
Every cloud model is reached through OpenRouter, one service that passes your question to the model you picked. When a cloud model answers, only your question and the rows it needs go, through OpenRouter, to the model company. They are not kept. With a local model, nothing leaves your computer.
Agents in the cloud (Pro)
On Pro, your agents can work while your devices are off. They run in the cloud, as part of Cloud Sync, and use the same one door. Free has no agents in the cloud, so on Free your agents run on your one device. They borrow tools (web search, later a browser or a sealed computer) only for a job.
- What they can read: Cloud Sync collections, if the AI setting allows it. End-to-end encrypted collections (only your devices hold the key) only with a key you give the agent. This device only, never.
- Every outside change (sending, posting, buying, changing something on a site) goes through one gate, the outbox. By default, what the agent starts on its own waits for your Approve tap.
- The AI model is never sent your password. 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.
How it is built: How agents in the cloud work (technical).