daggr is a Python library for building AI workflows that connect Gradio apps, ML models (through Hugging Face Inference Providers), and custom Python functions. It automatically generates a visual canvas for your workflow allowing you to inspect intermediate outputs, rerun any step any number of times, and preserves state for complex or long-running workflows. Daggr also tracks provenance: when you browse through previous results, it automatically restores the exact inputs that produced each output, and visually indicates which parts of your workflow are stale.

pip install daggr

(requires Python 3.10 or higher).

After installing daggr, create a new Python file, say app.py, and paste this code:

import random

import gradio as gr

from daggr import GradioNode, Graph

glm_image = GradioNode(
    "hf-applications/Z-Image-Turbo",
    api_name="/generate_image",
    inputs={
        "prompt": gr.Textbox(  # An input node is created for the prompt
            label="Prompt",
            value="A cheetah in the grassy savanna.",
            lines=3,
        ),
        "height": 1024,  # Fixed value (does not appear in the canvas)
        "width": 1024,  # Fixed value (does not appear in the canvas)
        "seed": random.random,  # Functions are rerun every time the workflow is run (not shown in the canvas)
    },
    outputs={
        "image": gr.Image(
            label="Image"  # Display original image
        ),
    },
)

background_remover = GradioNode(
    "hf-applications/background-removal",
    api_name="/image",
    inputs={
        "image": glm_image.image,
    },
    postprocess=lambda _, final: final,
    outputs={
        "image": gr.Image(label="Final Image"),  # Display only final image
    },
)

graph = Graph(
    name="Transparent Background Image Generator", nodes=[glm_image, background_remover]
)

graph.launch()

Run daggr app.py to start the app with hot reloading (or python app.py for standard execution). You should see a Daggr app like the one shown above that you can use to generate images with a transparent background!

Use Daggr when:

• You want to define an AI workflow in Python involving Gradio Spaces, inference providers, or custom functions
• The workflow is complex enough that inspecting intermediate outputs or rerunning individual steps is useful
• You need a fixed pipeline that you or others can run with different inputs
• You want to explore variations: generate multiple outputs, compare them, and always know exactly what inputs produced each result

Why not... ComfyUI? ComfyUI is a visual node editor where you build workflows by dragging and connecting nodes. Daggr takes a code-first approach: you define workflows in Python and the visual canvas is generated automatically. If you prefer writing code over visual editing, Daggr may be a better fit. In addition, Daggr works with Gradio Spaces and Hugging Face models directly, no need for specialized nodes.

Why not... Airflow/Prefect? Daggr was inspired by Airflow/Prefect, but whereas the focus of these orchestration platforms is scheduling, monitoring, and managing pipelines at scale, Daggr is built for interactive AI/ML workflows with real-time visual feedback and immediate execution, making it ideal for prototyping, demos, and workflows where you want to inspect intermediate outputs and rerun individual steps on the fly.

Why not... Gradio? Gradio creates web UIs for individual ML models and demos. While complex workflows can be built in Gradio, they often fail in ways that are hard to debug when using the Gradio app. Daggr tries to provide a transparent, easily-inspectable way to chain multiple Gradio apps, custom Python functions, and inference providers through a visual canvas.

Don't use Daggr when:

• You need a simple UI for a single model or function - consider using Gradio directly
• You want a node-based editor for building workflows visually - consider using ComfyUI instead

A Daggr workflow consists of nodes connected in a directed graph. Each node represents a computation: a Gradio Space API call, an inference call to a model, or a Python function.

Each node has input ports and output ports, which correspond to the node's parameters and return values. Ports are how data flows between nodes.

Input ports can be connected to:

• A previous node's output port → creates an edge, data flows automatically
• A Gradio component → creates a standalone input in the UI
• A fixed value → passed directly, doesn't appear in UI
• A Callable → called each time the node runs (useful for random seeds)

Output ports can be:

• A Gradio component → displays the output in the node's card
• None → output not displayed in the node's card but port can still connect to downstream nodes

Calls a Gradio Space API endpoint. Use this to connect to any Gradio app on Hugging Face Spaces or running locally.

from daggr import GradioNode
import gradio as gr

image_gen = GradioNode(
    space_or_url="black-forest-labs/FLUX.1-schnell",  # HF Space ID or URL
    api_name="/infer",                                 # API endpoint name
    inputs={
        "prompt": gr.Textbox(label="Prompt"),          # Creates UI input
        "seed": 42,                                    # Fixed value
        "width": 1024,
        "height": 1024,
    },
    outputs={
        "image": gr.Image(label="Generated Image"),   # Display in node card
    },
)

Finding the right inputs: To find what parameters a GradioNode expects, go to the Gradio Space and click "Use via API" at the bottom of the page. This shows you the API endpoints and their parameters. For example, if the API page shows:

from gradio_client import Client

client = Client("black-forest-labs/FLUX.1-schnell")
result = client.predict(
    prompt="Hello!!",
    seed=0,
    randomize_seed=True,
    width=1024,
    height=1024,
    num_inference_steps=4,
    api_name="/infer"
)

Then your GradioNode inputs should use the same parameter names: prompt, seed, randomize_seed, width, height, num_inference_steps.

Outputs: Output port names can be anything you choose—they simply map to the return values of the API endpoint in order. If an endpoint returns (image, seed), you might define:

outputs={
    "generated_image": gr.Image(),  # Maps to first return value
    "used_seed": gr.Number(),       # Maps to second return value
}

Runs a Python function. Input ports are automatically discovered from the function signature.

from daggr import FnNode
import gradio as gr

def summarize(text: str, max_words: int = 100) -> str:
    words = text.split()[:max_words]
    return " ".join(words) + "..."

summarizer = FnNode(
    fn=summarize,
    inputs={
        "text": gr.Textbox(label="Text to Summarize", lines=5),
        "max_words": gr.Slider(minimum=10, maximum=500, value=100, label="Max Words"),
    },
    outputs={
        "summary": gr.Textbox(label="Summary"),
    },
)

Inputs: Keys in the inputs dict must match the function's parameter names. If you don't specify an input, it uses the function's default value (if available).

Outputs: Return values are mapped to output ports in the same order they are defined in the outputs dict—just like GradioNode. For a single output, simply return the value. For multiple outputs, return a tuple:

def process(text: str) -> tuple[str, int]:
    return text.upper(), len(text)

node = FnNode(
    fn=process,
    inputs={"text": gr.Textbox()},
    outputs={
        "uppercase": gr.Textbox(),  # First return value
        "length": gr.Number(),       # Second return value
    },
)

Note: If you return a dict or list, it will be treated as a single value (mapped to the first output port), not as a mapping to output ports.

Concurrency: By default, FnNodes execute sequentially (one at a time per user session) to prevent resource contention from concurrent function calls. If your function is safe to run in parallel, you can enable concurrent execution:

# Allow this node to run in parallel with other nodes
node = FnNode(my_func, concurrent=True)

# Share a resource limit with other nodes (e.g., GPU memory)
gpu_node_1 = FnNode(process_image, concurrency_group="gpu", max_concurrent=2)
gpu_node_2 = FnNode(enhance_image, concurrency_group="gpu", max_concurrent=2)

Parameter	Default	Description
concurrent	False	If True, allow parallel execution
concurrency_group	None	Name of a group sharing a concurrency limit
max_concurrent	1	Max parallel executions in the group

Tip: When possible, prefer GradioNode or InferenceNode over FnNode. These nodes automatically run concurrently (they're external API calls), and your Hugging Face token is automatically passed through for ZeroGPU quota tracking, private Spaces access, and gated model access.

Calls a model via Hugging Face Inference Providers. This lets you use models hosted on the Hugging Face Hub without downloading them.

from daggr import InferenceNode
import gradio as gr

llm = InferenceNode(
    model="meta-llama/Llama-3.1-8B-Instruct",
    inputs={
        "prompt": gr.Textbox(label="Prompt", lines=3),
    },
    outputs={
        "response": gr.Textbox(label="Response"),
    },
)

Inputs: The expected inputs depend on the model's task type. For text generation models, use prompt. For other tasks, check the model's documentation on the Hub.

Outputs: Like other nodes, output names are arbitrary and map to return values in order.

Tip: InferenceNode and GradioNode automatically run concurrently and pass your HF token for ZeroGPU, private Spaces, and gated models. Prefer these over FnNode when possible.

GradioNode, FnNode, and InferenceNode all support optional preprocess and postprocess hooks that transform data on the way in and out of a node.

preprocess receives the input dict and returns a modified dict before the node executes. This is useful when an upstream node outputs data in a different format than the downstream node expects:

def fix_image_input(inputs):
    img = inputs.get("image")
    if isinstance(img, dict) and "path" in img:
        inputs["image"] = img["path"]
    return inputs

describer = GradioNode(
    "vikhyatk/moondream2",
    api_name="/answer_question",
    preprocess=fix_image_input,
    inputs={"image": image_gen.result, "prompt": "Describe this image."},
    outputs={"description": gr.Textbox()},
)

postprocess receives the raw return values from the node and lets you reshape them before they are mapped to output ports. If the node returns multiple values (a tuple), each value is passed as a separate argument. This is essential when working with Spaces that return extra values you don't need:

background_remover = GradioNode(
    "hf-applications/background-removal",
    api_name="/image",
    inputs={"image": some_node.image},
    postprocess=lambda original, final: final,  # Space returns (original, processed); keep only processed
    outputs={"image": gr.Image(label="Result")},
)

Another common pattern is extracting a specific item from a complex return value:

image_gen = GradioNode(
    "multimodalart/stable-cascade",
    api_name="/run",
    inputs={...},
    postprocess=lambda images, seed_used, seed_number: images[0]["image"],  # Extract first image
    outputs={"image": gr.Image(label="Generated Image")},
)

Key difference from Gradio: In daggr, all file-based data (images, audio, video, 3D models) is passed between nodes as file path strings. Gradio's type parameter (e.g., Image(type="numpy")) is ignored — daggr does not convert files to numpy arrays, PIL images, or any other in-memory format.

This means:

• Inputs to your node arrive as file path strings (e.g., "/tmp/daggr/abc123.png")
• Outputs from your node should be file path strings pointing to a file on disk

If your node expects a different format, use preprocess to convert file paths on the way in, and postprocess to convert back to file paths on the way out. This works with all node types:

from PIL import Image

def load_image(inputs):
    inputs["image"] = Image.open(inputs["image"])
    return inputs

def save_image(result):
    out_path = "/tmp/processed.png"
    result.save(out_path)
    return out_path

node = FnNode(
    lambda image: image.rotate(90),
    preprocess=load_image,
    postprocess=save_image,
    inputs={"image": gr.Image(label="Input")},
    outputs={"output": gr.Image(label="Rotated")},
)

For audio:

import soundfile as sf

def load_audio(inputs):
    data, sr = sf.read(inputs["audio"])
    inputs["audio"] = (sr, data)
    return inputs

def save_audio(result):
    sr, data = result
    out_path = "/tmp/processed.wav"
    sf.write(out_path, data, sr)
    return out_path

Different node types have different concurrency behaviors:

Node Type	Concurrency	Why
GradioNode	Concurrent	External API calls—safe to parallelize
InferenceNode	Concurrent	External API calls—safe to parallelize
FnNode	Sequential (default)	Local Python code may have resource constraints

Why sequential by default for FnNode? Local Python functions often:

• Access shared resources (files, databases, GPU memory)
• Use libraries that aren't thread-safe
• Consume significant CPU/memory

By running FnNodes sequentially per session, daggr prevents race conditions and resource contention. If your function is safe to run in parallel, opt in with concurrent=True.

Concurrency groups let multiple nodes share a resource limit:

# Both nodes share GPU—at most 2 concurrent executions total
upscale = FnNode(upscale_image, concurrency_group="gpu", max_concurrent=2)
enhance = FnNode(enhance_image, concurrency_group="gpu", max_concurrent=2)

You can test-run any node in isolation using the .test() method:

tts = GradioNode("mrfakename/MeloTTS", api_name="/synthesize", ...)
result = tts.test(text="Hello world", speaker="EN-US")
# Returns: {"audio": "/path/to/audio.wav"}

If called without arguments, .test() auto-generates example values using each input component's .example_value() method:

result = tts.test()  # Uses gr.Textbox().example_value(), etc.

This is useful for quickly checking what format a node returns without wiring up a full workflow.

Each node's inputs dict accepts four types of values:

Type	Example	Result
Gradio component	gr.Textbox(label="Topic")	Creates UI input
Port reference	other_node.output_name	Connects nodes
Fixed value	"Auto" or 42	Constant, no UI
Callable	random.random	Called each run, no UI

Each node's outputs dict accepts two types of values:

Type	Example	Result
Gradio component	gr.Image(label="Result")	Displays output in node card
None	None	Hidden, but can connect to downstream nodes

When a node outputs a list and you want to process each item individually, use .each to scatter and .all() to gather:

script = FnNode(fn=generate_script, inputs={...}, outputs={"lines": gr.JSON()})

tts = FnNode(
    fn=text_to_speech,
    inputs={
        "text": script.lines.each["text"],      # Scatter: run once per item
        "speaker": script.lines.each["speaker"],
    },
    outputs={"audio": gr.Audio()},
)

final = FnNode(
    fn=combine_audio,
    inputs={"audio_files": tts.audio.all()},    # Gather: collect all outputs
    outputs={"audio": gr.Audio()},
)

Sometimes you want to offer multiple alternatives for the same step in your workflow—for example, two different TTS providers or image generators. Use the | operator to create a choice node that lets users switch between variants in the UI:

host_voice = GradioNode(
    space_or_url="abidlabs/tts",
    api_name="/generate_voice_design",
    inputs={
        "voice_description": gr.Textbox(label="Host Voice"),
        "language": "Auto",
        "text": "Hi! I'm the host!",
    },
    outputs={"audio": gr.Audio(label="Host Voice")},
) | GradioNode(
    space_or_url="mrfakename/E2-F5-TTS",
    api_name="/basic_tts",
    inputs={
        "ref_audio_input": gr.Audio(label="Reference Audio"),
        "gen_text_input": gr.Textbox(label="Text to Generate"),
    },
    outputs={"audio": gr.Audio(label="Host Voice")},
)

# Downstream nodes connect to host_voice.audio regardless of which variant is selected
dialogue = FnNode(
    fn=generate_dialogue,
    inputs={"host_voice": host_voice.audio, ...},
    ...
)

In the canvas, choice nodes display an accordion UI where you can:

• See all available variants
• Click to select which variant to use
• View the selected variant's input components

The selected variant is persisted per sheet, so your choice is remembered across page refreshes. All variants must have the same output ports (so downstream connections work regardless of selection), but they can have different input ports.

import gradio as gr
from daggr import FnNode, GradioNode, Graph

# Generate voice profiles
host_voice = GradioNode(
    space_or_url="abidlabs/tts",
    api_name="/generate_voice_design",
    inputs={
        "voice_description": gr.Textbox(label="Host Voice", value="Deep British voice..."),
        "language": "Auto",
        "text": "Hi! I'm the host.",
    },
    outputs={"audio": gr.Audio(label="Host Voice")},
)

guest_voice = GradioNode(
    space_or_url="abidlabs/tts",
    api_name="/generate_voice_design",
    inputs={
        "voice_description": gr.Textbox(label="Guest Voice", value="Friendly American voice..."),
        "language": "Auto",
        "text": "Hi! I'm the guest.",
    },
    outputs={"audio": gr.Audio(label="Guest Voice")},
)

# Generate dialogue (would be an LLM call in production)
def generate_dialogue(topic: str, host_voice: str, guest_voice: str) -> tuple[list, str]:
    dialogue = [
        {"voice": host_voice, "text": "Hello, how are you?"},
        {"voice": guest_voice, "text": "I'm great, thanks!"},
    ]
    html = "<b>Host:</b> Hello!<br><b>Guest:</b> I'm great!"
    return dialogue, html  # Returns tuple: first value -> "json", second -> "html"

dialogue = FnNode(
    fn=generate_dialogue,
    inputs={
        "topic": gr.Textbox(label="Topic", value="AI"),
        "host_voice": host_voice.audio,
        "guest_voice": guest_voice.audio,
    },
    outputs={
        "json": gr.JSON(visible=False),  # Maps to first return value
        "html": gr.HTML(label="Script"),  # Maps to second return value
    },
)

# Generate audio for each line (scatter)
def text_to_speech(text: str, audio: str) -> str:
    return audio  # Would call TTS model in production

samples = FnNode(
    fn=text_to_speech,
    inputs={
        "text": dialogue.json.each["text"],
        "audio": dialogue.json.each["voice"],
    },
    outputs={"audio": gr.Audio(label="Sample")},
)

# Combine all audio (gather)
def combine_audio(audio_files: list[str]) -> str:
    from pydub import AudioSegment
    combined = AudioSegment.empty()
    for path in audio_files:
        combined += AudioSegment.from_file(path)
    combined.export("output.mp3", format="mp3")
    return "output.mp3"

final = FnNode(
    fn=combine_audio,
    inputs={"audio_files": samples.audio.all()},
    outputs={"audio": gr.Audio(label="Full Podcast")},
)

graph = Graph(name="Podcast Generator", nodes=[host_voice, guest_voice, dialogue, samples, final])
graph.launch()

Create a public URL to share your workflow with others:

graph.launch(share=True)

This generates a temporary public URL (expires in 1 week) using Gradio's tunneling infrastructure.

For permanent hosting, use daggr deploy to deploy your app to Hugging Face Spaces:

daggr deploy my_app.py

Assuming you are logged in locally with your Hugging Face token, this command:

1. Extracts the Graph from your script
2. Creates a Space named after your Graph (e.g., "Podcast Generator" → podcast-generator)
3. Uploads your script and dependencies
4. Configures the Space with the Gradio SDK

# Custom Space name
daggr deploy my_app.py --name my-custom-space

# Deploy to an organization
daggr deploy my_app.py --org huggingface

# Private Space with GPU
daggr deploy my_app.py --private --hardware t4-small

# Add secrets (e.g., API keys)
daggr deploy my_app.py --secret HF_TOKEN=xxx --secret OPENAI_KEY=yyy

# Preview without deploying
daggr deploy my_app.py --dry-run

Option	Short	Description
--name	-n	Space name (default: derived from Graph name)
--title	-t	Display title (default: Graph name)
--org	-o	Organization to deploy under
--private	-p	Make the Space private
--hardware		Hardware tier: cpu-basic, cpu-upgrade, t4-small, t4-medium, a10g-small, etc.
--secret	-s	Add secrets (repeatable)
--requirements	-r	Custom requirements.txt path
--dry-run		Preview what would be deployed

The deploy command automatically:

• Detects local Python imports and includes them
• Uses existing requirements.txt if present, or generates one with daggr
• Renames your script to app.py (HF Spaces convention)
• Generates the required README.md with Space metadata

You can also deploy manually by creating a new Space with the Gradio SDK, adding your workflow code to app.py, and including daggr in your requirements.txt.

Daggr automatically reads the GRADIO_SERVER_NAME and GRADIO_SERVER_PORT environment variables, which Hugging Face Spaces sets automatically for Gradio apps. This means your daggr app will work on Spaces without any additional configuration.

Daggr automatically saves your workflow state—input values, node results, and canvas position—so you can pick up where you left off after a page reload.

Sheets are like separate workspaces within a single Daggr app. Each sheet has its own:

• Input values for all nodes
• Cached results from previous runs
• Canvas zoom and pan position

Use sheets to work on multiple projects within the same workflow. For example, in a podcast generator app, each sheet could represent a different podcast episode you're working on.

The sheet selector appears in the title bar. Click to switch between sheets, create new ones, rename them (double-click), or delete them.

Every time a node runs, Daggr saves not just the output, but also a snapshot of all input values at that moment. This enables powerful exploratory workflows:

Browsing previous results: Use the ‹ and › arrows in the node footer to navigate through all cached results for that node (shown as "1/3", "2/3", etc.).

Automatic input restoration: When you select a previous result, Daggr automatically restores the input values that produced it. This means you can:

1. Generate multiple variations by running a node several times with different inputs
2. Browse through your results to find the best one
3. When you select a result, see exactly what inputs created it
4. Continue your workflow from that point with all the original context intact

Cascading restoration: When you toggle through results on a node, Daggr also automatically selects the matching result on downstream nodes (if one exists). For example, if you generated 3 images and removed the background from 2 of them, selecting image #1 will automatically show background-removal result #1.

Daggr uses edge colors to show you which parts of your workflow are up-to-date:

Edge Color	Meaning
Orange	Fresh—the downstream node ran with this exact upstream value
Gray	Stale—the upstream value has changed, or the downstream hasn't run yet

Edges are stale when:

• You edit an input value (e.g., change a text prompt)
• You select a different cached result on an upstream node
• A downstream node hasn't been run yet

This visual feedback helps you understand at a glance which results are current and which need to be re-run. It's especially useful in long workflows where you might forget which steps you've already executed with your current inputs.

Example workflow:

1. Generate an image with prompt "A cheetah in the savanna" → edge turns orange
2. Edit the prompt to "A lion in the jungle" → edge turns gray (stale)
3. Re-run the image generation → edge turns orange again
4. Run the background removal node → that edge also turns orange

This provenance tracking is particularly valuable for creative workflows where you're exploring variations and want to always know exactly what inputs produced each output.

Environment	User Status	Persistence
Local	Not logged in	✅ Saved as "local" user
Local	HF logged in	✅ Saved under your HF username

When running locally, your data is stored in a SQLite database at ~/.cache/huggingface/daggr/sessions.db.

By default, the persist_key is derived from your graph's name:

Graph(name="My Podcast Generator")  # persist_key = "my_podcast_generator"

If you later rename your app but want to keep the existing saved data, set persist_key explicitly:

Graph(name="Podcast Generator v2", persist_key="my_podcast_generator")

For scratch workflows or demos where you don't want data saved:

Graph(name="Quick Demo", persist_key=False)

This disables all persistence—no sheets UI, no saved state.

Daggr automatically uses your local Hugging Face token for both GradioNode and InferenceNode. This enables:

• ZeroGPU quota tracking: Your HF token is sent to Gradio Spaces running on ZeroGPU, so your usage is tracked against your account's quota
• Private Spaces access: Connect to private Gradio Spaces you have access to
• Gated models: Use gated models on Hugging Face that require accepting terms of service

To log in with your Hugging Face account:

pip install huggingface_hub
hf auth login

You'll be prompted to enter your token, which you can find at https://huggingface.co/settings/tokens.

Once logged in, the token is saved locally and daggr will automatically use it for all GradioNode and InferenceNode calls—no additional configuration needed.

Alternatively, you can set the HF_TOKEN environment variable directly:

export HF_TOKEN=hf_xxxxx

Daggr is designed to be LLM-friendly, making it easy for AI coding assistants to generate and debug workflows. To give your AI coding assistant context on how to use daggr, you can install the daggr skill:

npx skills add gradio-app/daggr

When you (or an LLM) make a mistake, Daggr provides detailed, actionable error messages with suggestions:

Invalid API endpoint:

ValueError: API endpoint '/infer' not found in 'hf-applications/background-removal'. 
Available endpoints: ['/image', '/text', '/png']. Did you mean '/image'?

Typo in parameter name:

ValueError: Invalid parameter(s) {'promt'} for endpoint '/generate_image' in 
'hf-applications/Z-Image-Turbo'. Did you mean: 'promt' -> 'prompt'? 
Valid parameters: {'width', 'height', 'seed', 'prompt'}

Missing required parameter:

ValueError: Missing required parameter(s) {'prompt'} for endpoint '/generate_image' 
in 'hf-applications/Z-Image-Turbo'. These parameters have no default values.

Invalid output port reference:

ValueError: Output port 'img' not found on node 'Z-Image-Turbo'. 
Available outputs: image. Did you mean 'image'?

Invalid function parameter:

ValueError: Invalid input(s) {'toppic'} for function 'generate_dialogue'. 
Did you mean: 'toppic' -> 'topic'? Valid parameters: {'topic', 'host_voice', 'guest_voice'}

Invalid model name:

ValueError: Model 'meta-llama/nonexistent-model' not found on Hugging Face Hub. 
Please check the model name is correct (format: 'username/model-name').

These errors make it easy for LLMs to understand what went wrong and fix the generated code automatically, enabling a smoother AI-assisted development experience.

When building workflows, LLMs can use .test() to discover a node's actual output format:

# LLM wants to understand what whisper returns
whisper = InferenceNode("openai/whisper-large-v3", inputs={"audio": gr.Audio()})
result = whisper.test(audio="sample.wav")
# Returns: {"text": "Hello, how are you?"}

This helps LLMs:

• Understand the structure of node outputs
• Apply postprocess functions to extract specific values
• Create intermediate FnNodes to transform data between nodes

For example, if a node returns multiple values but you only need one:

# After discovering the output format with .test()
bg_remover = GradioNode(
    "hf-applications/background-removal",
    api_name="/image",
    inputs={"image": some_image.output},
    postprocess=lambda original, final: final,  # Keep only the second output
    outputs={"image": gr.Image()},
)

While in our examples above, we've seen how Daggr works with remote Gradio Spaces and Hugging Face Inference Providers, it's also well-suited for completely local, offline workflows.

The easiest way to run a Space locally is to set run_locally=True on any GradioNode. Daggr will automatically clone the Space, install dependencies in an isolated virtual environment, and launch the Gradio app:

from daggr import GradioNode, Graph
import gradio as gr

# Automatically clone and run the Space locally
background_remover = GradioNode(
    "hf-applications/background-removal",
    api_name="/image",
    run_locally=True,  # Run locally instead of calling the remote API
    inputs={"image": gr.Image(label="Input Image")},
    outputs={"final_image": gr.Image(label="Output")},
)

graph = Graph(name="Local Background Removal", nodes=[background_remover])
graph.launch()

On first run, daggr will:

1. Clone the Space repository to ~/.cache/huggingface/daggr/spaces/
2. Create an isolated virtual environment with the Space's dependencies
3. Launch the Gradio app on an available port
4. Connect to it automatically

Subsequent runs reuse the cached clone and venv, making startup much faster.

If local execution fails (missing dependencies, GPU requirements, etc.), daggr automatically falls back to the remote API and prints helpful guidance:

⚠️  Local execution failed for 'owner/space-name'
Reason: Failed to install dependencies
Logs: ~/.cache/huggingface/daggr/logs/owner_space-name_pip_install_2026-01-27.log
Falling back to remote API...

To disable fallback and see the full error (useful for debugging):

export DAGGR_LOCAL_NO_FALLBACK=1

Variable	Default	Description
DAGGR_LOCAL_TIMEOUT	120	Seconds to wait for the app to start
DAGGR_LOCAL_VERBOSE	0	Set to 1 to show app stdout/stderr
DAGGR_LOCAL_NO_FALLBACK	0	Set to 1 to disable fallback to remote
DAGGR_UPDATE_SPACES	0	Set to 1 to re-clone cached Spaces
DAGGR_DEPENDENCY_CHECK	(unset)	skip, update, or error — controls upstream hash checking
GRADIO_SERVER_NAME	127.0.0.1	Host to bind to. Set to 0.0.0.0 on HF Spaces
GRADIO_SERVER_PORT	7860	Port to bind to

You can also run a Gradio app yourself and point to it directly:

from daggr import GradioNode, Graph
import gradio as gr

# Connect to a Gradio app you're running locally
local_model = GradioNode(
    "http://localhost:7860",  # Local URL instead of Space ID
    api_name="/predict",
    inputs={"text": gr.Textbox(label="Input")},
    outputs={"result": gr.Textbox(label="Output")},
)

graph = Graph(name="Local Workflow", nodes=[local_model])
graph.launch()

This approach lets you run your entire workflow offline, use custom or fine-tuned models, and avoid API rate limits.

Daggr workflows can be called programmatically via REST API, making it easy to integrate workflows into other applications or run automated tests.

First, get the API schema to see available inputs and outputs:

curl http://localhost:7860/api/schema

Response:

{
  "subgraphs": [
    {
      "id": "main",
      "inputs": [
        {"node": "image_gen", "port": "prompt", "type": "textbox", "id": "image_gen__prompt"}
      ],
      "outputs": [
        {"node": "background_remover", "port": "image", "type": "image"}
      ]
    }
  ]
}

Execute the entire workflow by POSTing inputs to /api/call:

curl -X POST http://localhost:7860/api/call \
  -H "Content-Type: application/json" \
  -d '{"inputs": {"image_gen__prompt": "A mountain landscape"}}'

Response:

{
  "outputs": {
    "background_remover": {
      "image": "/file/path/to/output.png"
    }
  }
}

Input keys follow the format {node_name}__{port_name} (with spaces/dashes replaced by underscores).

If your workflow has multiple disconnected subgraphs, use /api/call/{subgraph_id}:

# List available subgraphs
curl http://localhost:7860/api/schema

# Call a specific subgraph
curl -X POST http://localhost:7860/api/call/subgraph_0 \
  -H "Content-Type: application/json" \
  -d '{"inputs": {...}}'

import requests

# Get schema
schema = requests.get("http://localhost:7860/api/schema").json()

# Execute workflow
response = requests.post(
    "http://localhost:7860/api/call",
    json={"inputs": {"my_node__text": "Hello world"}}
)
outputs = response.json()["outputs"]

During development, you can use the daggr CLI to run your app with automatic hot reloading. When you make changes to your Python file or its dependencies, the app automatically restarts:

daggr examples/01_quickstart.py

This is much faster than manually stopping and restarting your app each time you make a change.

daggr <script> [options]

Option	Description
--host	Host to bind to (default: 127.0.0.1)
--port	Port to bind to (default: 7860)
--no-reload	Disable auto-reload
--no-watch-daggr	Don't watch daggr source for changes

By default, the CLI watches for changes in:

• Your script file and its directory
• Local imports from your script
• The daggr source code itself (useful when developing daggr)

To disable watching the daggr source (e.g., in production-like testing):

daggr examples/01_quickstart.py --no-watch-daggr

To speed up reloads, daggr caches Gradio Space API info in ~/.cache/huggingface/daggr/. This means:

• First run: Connects to each Gradio Space to fetch API info (cached to disk)
• Subsequent reloads: Loads from cache, no network calls needed

If you change a Space's API or encounter stale cache issues, clear the cache:

rm -rf ~/.cache/huggingface/daggr

Use daggr <script> when you're actively developing and want instant feedback on changes.

Use python <script> when you want the standard behavior (no file watching, direct execution).

When your workflow references external Gradio Spaces or Hugging Face models, those dependencies can change at any time—a Space author might update the model, change the API, or alter default behavior. This can silently break reproducibility: the same workflow with the same inputs may produce different results weeks later.

To address this, daggr tracks the commit SHA of every upstream Space and model the first time your app launches. On subsequent launches, daggr compares the cached SHA against the current version. If an upstream dependency has changed, you'll see a terminal warning:

  ⚠️  Upstream dependency changes detected:

    • space 'mrfakename/MeloTTS' (node: MeloTTS)
      cached:  a1b2c3d4e5f6
      current: f6e5d4c3b2a1

  How would you like to handle 'mrfakename/MeloTTS'?
    [1] Duplicate the original version under your namespace (safer)
    [2] Update to the latest version

Option 1 (Spaces only) downloads the Space at the exact commit you originally built against and re-uploads it under your Hugging Face namespace, so your workflow continues using the known-good version. This requires being logged in via huggingface-cli login.

Option 2 updates the cached hash to accept the new version.

For CI/CD or non-interactive environments, set the DAGGR_DEPENDENCY_CHECK environment variable:

Value	Behavior
skip	Skip all dependency checks
update	Auto-accept upstream changes
error	Fail if any dependency has changed

Dependency hashes are stored in ~/.cache/huggingface/daggr/_dependency_hashes.json.

pip install -e ".[dev]"
ruff check --fix --select I && ruff format

MIT License