Creative AI is evolving from single-step asset generation to long-horizon multimodal production, but existing tools—prompt-based, chat-based, or node-based—fall short in supporting the complex, evolving state of real-world creative projects. They often discard intermediate context, rely on linear conversations, or require manual workflow specification. To bridge this gap, Hugging Face researchers have introduced JarvisHub, an open-source harness for canvas-native creative agents.
JarvisHub treats an editable canvas as the user workspace, the agent's external memory, action space, and shared project state. Multimodal artifacts, dependencies, versions, and feedback are represented as typed canvas nodes and links. A three-layer architecture—canvas state, protocol bridge, and agent runtime—enables agents to act within an inspectable and editable creative state, moving beyond isolated tool use toward sustained, human-steerable creative automation.
Key features include: Canvas as Memory—prompts, references, candidate versions, outputs, and feedback remain on the canvas, allowing the agent to see the entire project and continuously advance it; Inspectable and Recoverable—generation process, dependencies, and revision history are visible, and errors can be fixed by editing the relevant node without restarting; and Long-Horizon Multimodal Creation—native support for images, videos, websites, presentations, and more.
JarvisHub currently showcases three representative use cases: narrative media generation, interactive web development, and presentation creation. The project is now live and open source, inviting the community to explore and contribute.