I built a canvas you and an AI agent can both edit

I’m building Galta, a canvas that opens beside the webpage you’re browsing. Drag in a reference, sketch an idea, and ask an agent to work on it. You can edit the same canvas yourself and keep going from there.

A casual test surprised me: GPT-6-Astra redrew a suit-cat meme in 620 editable vector strokes. Then I asked it to do it again without lifting the pen. The result was one continuous stroke with 70,722 points.

Demo: https://x.com/Koho_Z/status/2115504320712171614

The video is an edited, sped-up replay of the actual strokes.

I’d like to find a useful next test. When working with AI on something visual, which part do you keep having to describe in words when you wish you could just point at it?

Try it: https://galta.art

I’m the developer. Connecting AI currently requires an Apple silicon Mac; Codex and Claude Code connect through MCP.