It's interesting to watch projects like OpenDesign emerge almost in real time around the behaviors pioneered by tools like Claude Code, Codex and others. Not just from the "open-source replica" angle—even though the quality is mind-blowing—but rather through how much it deepens a strong architectural shift taking shape across multiple use cases:
Local-first runtimes, BYOK orchestration, filesystem-native memory, harness engineering, skills and design systems as portable text artifacts.
We see more and more AI agents dissolving the boundary between application, workspace and workflow.
What stands out in such projects—and their mother-ecosystem Nexu, originally started as a desktop client for OpenClaw—is that they do not merely propose another AI app, but an alternative, comprehensive theory of the agentic workspace itself: open-ended, distributed, model-agnostic, runtime-centric.
Seems like, more and more, the harness is the product.
Looking at OpenDesign and its counterparts, I also notice a fascinating asymmetry: the input layer is becoming almost paradoxically simple—a command, a sentence, an intent. The output layer is moving in the exact opposite direction: toward increasingly structured, visually rich artifacts, with a level of rigor that would have required entire design teams not long ago. See "html-anything," also by Nexu, or "hyperframes" (Heygen)—CLIs that generate beautifully designed, export-ready content directly.
The complexity hasn't vanished. It has migrated.
Behind OpenDesign and Nexu, we may glimpse the early contours of a broader ecosystem quietly taking shape within the Chinese OSS community.
Still extremely early, of course, and waiting to be battle-tested—but historically, paradigms often begin precisely like this: as strange-looking side projects, half community, half architecture manifesto.
