Daggr: A New Open-Source Python Library for Inspectable AI Workflows

The Gradio team has released Daggr, a new open-source Python library designed to simplify the construction and debugging of multi-step AI workflows. Daggr allows developers to define workflows programmatically in Python while automatically generating a visual canvas that exposes intermediate states, inputs, and outputs for each step in the pipeline.

The Problem Daggr Solves

Applied AI development often involves complex, multi-step pipelines where failures can occur at any stage. Traditional approaches force developers to rerun entire workflows after a failure, even when the issue occurs late in the process. This creates a slow feedback loop that hampers experimentation and debugging.

Daggr addresses this by organizing workflows as directed graphs, allowing for independent inspection and re-execution of each node. This approach enhances debugging and speeds up iteration by tackling the issue of slow and unclear experimentation.

Code-First Approach with Visual Feedback

Unlike GUI-driven workflow builders that often sacrifice version control and flexibility, Daggr follows a code-first philosophy. Developers define nodes and connections directly in Python, and Daggr renders a corresponding visual interface for inspection. This means the visual layer is derived from code rather than replacing it, allowing workflows to remain reproducible and easy to review.

Core Node Types

Daggr supports three primary node types that cover most AI workflow scenarios:

• GradioNode: Connects directly to Gradio apps or Hugging Face Spaces, enabling existing demos and tools to be reused as workflow components
• FnNode: Wraps arbitrary Python functions, making it possible to insert custom preprocessing or postprocessing logic
• InferenceNode: Interfaces with models served through Hugging Face Inference Providers, allowing hosted models to be incorporated without additional adapters

State Persistence and Debugging

A key feature is state persistence. Daggr automatically saves workflow state, cached results, input values, and canvas layout, allowing developers to pause and resume work without losing context. Individual nodes can be rerun with modified inputs, which is particularly useful when debugging long pipelines or comparing alternative implementations of a single step.

Integration with the Gradio Ecosystem

Because Daggr is built by the Gradio team, it integrates closely with the Gradio ecosystem. Workflows can be launched locally with a visual canvas served in the browser, or shared via public links using Gradio's tunneling. For longer-lived deployments, the same workflows can be hosted on Hugging Face Spaces by adding Daggr as a dependency.

Early Developer Reactions

Early reactions from developers have focused on the combination of programmatic control and visual feedback. Sebastian Buzdugan commented: "Mixing endpoints and Gradio is such a smart combination." Others have highlighted Daggr's usefulness for rapid experimentation.

Current Status and Availability

Daggr is a lightweight, experimental project currently in beta. Its API may change as users interact with it. While workflow state is stored locally, updates could still result in data loss, highlighting its purpose as a development and prototyping tool rather than a production solution.

The library is available today as an open-source Python package and can be installed via pip or uv for Python 3.10 and newer. The source code, examples, and documentation are published on GitHub, and the team has invited feedback and contributions as the library matures.

For developers building complex AI workflows, Daggr offers a compelling middle ground between pure code and visual builders, combining the reproducibility of programmatic workflows with the debugging power of visual inspection.