Brood is a macOS desktop app for iterating on visual ideas where the canvas image is the primary state and text is optional. The UI is an RTS-style tool palette: you pick an operation and apply it to regions of the canvas.

Under the hood, Brood routes each step to an image model/tool (Gemini, OpenAI, Flux) and runs a lightweight policy step that looks at the current canvas and chooses the next operation (eg background removal, recast/style shift, object replace) based on what it infers youre trying to do.

This started as an experiment inspired by Karpathys "the tokenizer must go" / image-input-first argument (https://x.com/karpathy/status/1980397031542989305 ).

Feedback I want:

Does the RTS-style palette speed up iteration or add distraction?

Which 2-3 operations should be next?