Introducing Daggr: Chain apps programmatically, inspect visually

Source: HuggingFace Date: 2026-01-29 URL: https://huggingface.co/blog/daggr

Summary

Library release (beta): Daggr, a Python library for building AI pipelines with automatic visual canvas generation for inspection. Code-first (version-controllable) but produces a visual DAG view; intermediate step outputs are inspectable and individually rerunnable. Three node types: GradioNode (calls Gradio Spaces/APIs), FnNode (custom Python), InferenceNode (HF Inference Providers). Native Gradio integration, state persistence. Demo: image-to-3D-asset pipeline.

Implications

Thread: HF as open-source ML hub / agentic patterns. Daggr occupies the same space as Langchain flows, Prefect, and similar pipeline tools but is HF-native — GradioNode and InferenceNode are first-class abstractions that assume HF as the execution environment. The visual inspection without visual authoring (code defines the workflow, canvas is for debugging) is a pragmatic design choice that avoids the YAML-trap. Beta APIs will change. Watch whether this becomes the recommended way to chain HF Spaces, which would give it a natural moat over general pipeline tools.