Show HN: Fantastic Canvas
github.comFantastic is my own harness. It runs on an Erlang-like, OS-like meta-runtime instead of agent loops or agent graphs.
I believe this is the right design for an AI runtime: it puts humans, AI agents and tools on equal footing.
It will grow into a convenient multi-OS runtime that can do what other AI libraries and harnesses do. For now it’s a technical demo with two modes: a CLI coder, and a spatial canvas that’s very handy for a lot of everyday work.
I think the canvas mode is the kind of software we’ll all be using in the 2030s.
I don’t have a name for it yet, so for now let’s call it “emergent software”. The idea is that you build, run and engineer your own systems visually, with help from AI, inside your own dynamic interface that you build for yourself. It’s a lot of fun, especially for multimedia projects.
For example, in a canvas project I’m still working on, a third-party AI agent generates UI and agents: it builds a control panel for making a Manim animation, and a viewport for a Blender video built from the Manim animation’s guidelines. I’ll record a video later; for now I just want to share the general idea. Maybe someone will like it.
If anyone wants to talk about it more, let me know in the comments and I’ll respond.
Have a fantastic day!
Oleksandr
Hi, I'm interested in this but I don't understand how it actually works or how I could put it to use. Could you sort of dumb it down and explain it to me if you don't mind.
The easiest way is to build it locally (ask your AI agent to do it) and run it in "fantastic canvas" mode. That starts a local web app where you can add prebundled agents, and your CLI coder should be in the double-click menu. Start it, ask it to look around, then ask it to build something for you.
Most AI coders produce source code and a chat log. This one builds an interactive UI for you, which feels more like a personal garage where you build your artwork than the usual task list and chat log we've seen in Lovable or the agentic VS Code view.
You can also ask your AI agent how to use the kernel to build your own agentic system. It may blow your mind: all actors live on one flat surface and exchange calls, with permissions on every call between them. It's harder to understand at first, but once you get it, you'll forget about agent loops and agent graphs and use it as the base runtime for your AI software.
I packed a lot in here, so it's not easy to explain; I hope this makes sense. In one sentence: a flat runtime that replaces agent loops and agent graphs, plus a canvas web app for your state-of-the-art AI workflow in the browser.