Assistant

How the assistant behaves and how the app looks. Keys and models aren't set up here; they are in the Models subsection.

Personal instructions

Custom instructions: what the assistant should know about how to answer you. You can also change them from a chat: say, for example, "answer me more briefly", and the assistant suggests new text with a card; you apply it, and you can undo it. For example: "answer briefly, explain terms in plain words, answer in English". The instructions are stored encrypted and apply to all new answers. They don't override MindDuck's privacy and safety rules.

Click Save preferences to apply the changes.

• Let the assistant suggest facts: allow the assistant to suggest facts to remember. Suggestions aren't used until you confirm them. See Facts.
• Search memory for every text message: an experimental mode: search memory before every message, not only when the question hints at it.
• Smart memory selection: before answering, the Jev model on OpenRouter looks through up to 300 fragments found in memory and keeps up to five that really help answer. The fragments go to OpenRouter with your key; an answer costs a fraction of a cent. If the selection doesn't fit into a second or fails, the regular one is used. You need an OpenRouter key in Models. Modes: Automatic (default) for chats through OpenRouter; Always for chats through other providers too; Off. If a chat goes through another provider and you have an OpenRouter key, MindDuck asks once in the chat whether to turn on the selection. Below the mode are Connection (which OpenRouter connection to use; Automatic takes the chat's connection or the first one with a key) and Relevance model (empty means the default model, currently typesafe/jev-1.13). The Test button checks the connection and the model on two short made-up fragments. You can ask the assistant to turn it on or off, and it will suggest the change with a card.

Size and spacing

Interface and text size, independently of each other. Applies only to this device: you can choose different sizes on your phone and your computer.

• Interface scale: the whole app: text, spacing, buttons;
• Chat text: only the text of messages, on top of the interface scale;
• Chat line spacing: the distance between lines in messages.
• Document font: the font of documents for reading and editing: Sans (Inter), Serif (Literata, a book font for long reading on screen), Mono or System (the device's font);
• Document text and Document line spacing: text size and line spacing in documents;
• Chat width and Document width: the width of the text column on large screens. You can also change it with the mouse: hover over the edge of the text in a chat or a document, a thin line appears, drag it. A double click on the edge returns the default width. While dragging you see roughly how many characters fit in a line; 65 to 80 is comfortable to read.

Web page reader

How the assistant opens pages from the internet: Local only (default), Automatic: this device first, then Jina Reader or Always use Jina Reader, and whether pages built by scripts open in the background browser on the device. The same place has the Local servers the assistant may open list: local addresses such as http://localhost:3000 the assistant may open. The list is stored only on this device. See Web search and reading pages.

Notifications

Notify when a response is ready or needs your permission: a system notification while MindDuck is in the background, when an answer is ready, when it fails, or when it waits for your permission (an MCP tool card or a request to look at a page). The notification names the chat and never shows the answer; a private chat is not named. Stopping an answer is not announced. The setting is stored on this device and is on by default.

Clicking a notification on a computer doesn't open the chat. On a phone the system may pause MindDuck in the background, so a long answer may finish only after you return.

Background

Computers only. Applies to this device.

• Keep MindDuck running when its window is closed - closing the window hides it, and MindDuck keeps running: sync, indexing and the import of agent sessions go on. On by default. The window comes back with a click on the MindDuck icon in the tray, or by starting MindDuck as usual; on a Mac, from the Dock. The first time the window is closed this way, MindDuck tells you once with a notification. With the setting off, closing the window quits MindDuck.
• Start MindDuck when you sign in to this computer - the system starts MindDuck at sign-in, without its window. Off by default.
• Quit MindDuck quits the app whatever the first setting says.

The tray icon menu (right click) has Open MindDuck, Lock workspace, which closes the workspace like Close workspace in the app, and Quit MindDuck. A Mac has no icon in the menu bar: the window opens from the Dock and Command-Q quits. On Linux the icon appears if the desktop environment supports one; without it, start MindDuck again to bring the window back.

A second copy of MindDuck does not start: starting it again shows the window of the one already running.

Model work display

What to show while the model answers: Detailed, Compact or Developer. It changes only the display. See How to read the answer.

Developer mode

Shows request traces, diagnostics and manual control of summaries and topic analysis. Not needed for regular work.

Under each assistant answer, the Developer trace row shows where the time before the first word went: Request prepared in … is building the request on this device (context, memory search), API answered after … is waiting for the provider's response. The same values are in the Export debug trace file, in the timing field (preparationMs, waitForFirstOutputMs). Answers created by earlier versions don't have this data.

What moved

• Sync and periodic Google Drive checks are in Sync & backup. The Check Google Drive while this app is open setting checks for changes about every 30 seconds after activity and every 5 minutes while idle.
• The background tasks mode (Run tasks) is in Models → Service.
• Fallback web search is in Models → Web search.