Inkmind

Every AI product is a chat box. inkmind is the counter-position, in running code: the interface is the stroke. You draw on an infinite whiteboard — a math step, a diagram, a half-formed question — and the AI draws its answer back, by hand, right where you’re working: circling a mistake, labelling the leg of a triangle, writing the next line of the proof. No prompt to type, no thread to scroll. The only channel between you and the model is ink.

Under the hood it’s deliberately un-scaffolded: one mind does everything. The whole board — your marks as an image, every element’s box, the running back-and-forth — goes to a single model that reads it, reasons about it, and emits its reply as a raw stream of pen points it places itself. No font, no shape library, no second model filling in the strokes. If it knows where the chess pieces go, it draws them there; if it’s solving your equation, it writes the steps in its own hand. The reasoning and the hand are the same act.

The strokes stream as they generate — each one parsed the moment it closes and drawn onto the board live, so you watch it think with a pen instead of staring at a spinner. It’s built on a real Excalidraw canvas (the full infinite whiteboard, every tool), so the AI’s marks land as elements you can move, keep, or redraw. It runs on the Anthropic API in production; a local mode can even draw from a Claude subscription instead of API credits.

The surprise the lab found building it: a model’s ability to draw is a different axis from its ability to reason. The flagship reasoner sets the problem up perfectly but draws a rough, half-legible hand; a model treated as lesser draws clean, readable strokes — a neutral reader can decode one and not the other. Nothing on the benchmarks measures spatial-motor generation, so the gap is invisible to the leaderboards and only shows up when you build something that has to actually draw. So inkmind defaults to the better hand, and reaches for the bigger brain only when a problem needs it.