Answers, fixes, and a direct line.

That's a suggestion from the built-in Smarti Assistant. It's dashed, muted, and a different hue on purpose: it lives in a separate layer from your board's data, so a suggestion cannot merge into your content — even by accident.

Accept or Dismiss, both one click. Accepting constructs a brand-new card with a small authorship dot; dismissing is remembered for the session, so it won't re-offer the same idea reworded.

Never per keystroke. It waits about four seconds of quiet after your last meaningful change, needs at least three non-empty ideas on the board, and offers one suggestion at a time. Dragging cards around never triggers it — only text, connections, the objective, and crossing ideas off do, because those change what the board says.

The wait is adjustable: ⚙ Settings → Ghost frequency offers 4s (default), 10s, 30s, 1 minute, or Off. Off stops the unsolicited ghost entirely; the ideas panel keeps working.

No account, ever — and no key is a supported configuration: the board is fully usable and the assistant simply stays off.

If you want the assistant, bring your own key: Anthropic (Claude), z.ai (GLM), z.ai Coding Plan, or any OpenAI-compatible endpoint (LM Studio, vLLM, OpenRouter…). Ollama runs locally with nothing leaving your machine. Keys are stored in your own database and are write-only — the settings panel only ever shows the last four characters.

⌘⇧P (or the Private button) marks the board you're looking at private: no suggestions arrive, the ideas button is unavailable, and nothing from that board is sent to any model.

It's per board, so the rest keep working. The promise is kept by the server, not the browser — a stale tab or a retry gets the same refusal — and ⌘Z can never switch it back on for you. Turning it off doesn't immediately wake the assistant; the next real edit does.

One SQLite file — ./data/smarti.db by default, one row per board. No accounts, no cloud, no telemetry. Back up by copying the file; move to another machine by copying the folder.

No. The assistant never moves, edits, or deletes a node you placed — that rule has no exception. The only bridge from a suggestion to the board is your click.

⌘Z undoes, ⌘⇧Z or ⌘Y redoes. Every deliberate change to content — typing, adding or deleting cards, accepting a suggestion, crossing an idea off, the objective — is one undo step. Dragging and resizing aren't content, so they don't clutter the stack.

A suggestion appearing is never in your undo stack; accepting one is a normal, reversible edit.

There is no save button — the board autosaves continuously, and the status in the bottom-right corner says so. Boards name themselves after their first idea; rename from the top-right corner of the canvas and the override sticks.

The first board in a fresh install is a tutorial — a real board you can edit, wreck, or delete. Open the tutorial board on the library page brings a fresh copy back any time.

⌘⇧F presents: fullscreen, read-only, the whole board fitted to view — pan and zoom still work, and Esc brings you back exactly where you were.

⌘P prints a read-only paper copy (or saves it as PDF) — the whole board on one page, shrinking as far as ~40% to fit; only a truly sprawling board tiles across pages. Done strikes and formatting print; the pending suggestion, selections, and search highlights never do.

Yes — the canvas answers touch: drag to pan, pinch to zoom, drag a card to move it, and press and hold where you would press Shift (empty canvas to sweep a selection, a card to add it to one). Double-tap for a new idea; double-tap a card to edit.

To reach the board from another device, start it with ./start.sh --lan and type the address it prints. That's opt-in on every run, and deliberately so: there's no login, so only do it on networks you trust.

Hand control (new in v6) turns your webcam into a navigator: pinch and move to pan, pinch with both hands and spread them to zoom, and holding a pinch near the edge keeps the board scrolling. It works on the canvas and in Present mode — which is where it shines.

The tracker is MediaPipe's hand-landmark model running inside the app itself — no cloud, no upload, and nothing is ever recorded. The camera is off until you press Hand control and stops the moment you turn it off. And it's not the assistant: it's local computer vision, never a token.

Pan and zoom — and that's the whole vocabulary. Hands can't select, edit, move, or cross off cards, and they can't accept or dismiss suggestions. Navigation-only is deliberate: your board's content stays out of the gesture vocabulary.

On-screen rings always show what the tracker sees — pinch state, zoom mode, and when you're close to losing it at the edge of the frame — and sensitivity is adjustable in ⚙ Settings.

No multiplayer, no freehand drawing or images, no cross-session memory, no cloud sync. One job, done well: thinking by arranging ideas — with an assistant that respects the arrangement.