Show HN: Comfy Nodekit – build/serialize ComfyUI workflows in Python

13 hours ago 1

Utilities for working with ComfyUI node graphs from Python. The project provides typed wrappers around ComfyUI nodes, lightweight graph primitives, and helpers for exporting a workflow back to the JSON format that the ComfyUI API expects.

Typed Node and Edge models (Pydantic) for tracking dependencies and serialising graphs.
A generated nodes.py module that mirrors the ComfyUI node catalogue with Python callables.
Tooling (nodes_to_py.py) to regenerate those bindings directly from a running ComfyUI server.
Example workflow builder scaffolding that shows how to stitch nodes into higher-level pipelines.

Python 3.12+
requests and pydantic (installed automatically when the package is installed)
Access to a running ComfyUI server when regenerating node bindings

uv add comfy_nodekit # pip install comfy_nodekit

Generate node bindings (required before usage)

Before using the factory functions, generate nodes.py from your running ComfyUI:

uv run python -m comfy_nodekit.nodes_to_py http://127.0.0.1:8188/ # or: just run

If you omit the URL, it defaults to http://127.0.0.1:8188/. This writes nodes.py based on the server's /object_info. Re-run whenever your ComfyUI nodes change.

You can also generate from a local JSON schema (previously fetched from /object_info):

# From default server (same as before) uv run python -m comfy_nodekit.nodes_to_py # From a specific server and write to a custom path uv run python -m comfy_nodekit.nodes_to_py http://127.0.0.1:8188/ ./nodes.py # From a local JSON file and write to a custom path uv run python -m comfy_nodekit.nodes_to_py ./object_info.json ./nodes.py

Create nodes the same way you would wire a ComfyUI graph. Each factory function from nodes.py returns a typed Edge that keeps track of its originating node. When you are ready to export the workflow, grab the source_node of your terminal edge and call build_json_workflow().

from comfy_nodekit.models import Edge from comfy_nodekit.nodes import ( CheckpointLoaderSimple, CLIPTextEncode, EmptyLatentImage, KSampler, VAEDecode, ) # Load base models model, clip, vae = CheckpointLoaderSimple("juggernautXL_v9Rdphoto2Lightning.safetensors") # Encode prompts positive = CLIPTextEncode(text="a cinematic portrait of an astronaut", clip=clip) negative = CLIPTextEncode(text="", clip=clip) # Create latents and sample latent = EmptyLatentImage(width=1024, height=1024, batch_size=1) sampled = KSampler( model=model, seed=42, steps=20, cfg=7.5, sampler_name="euler", scheduler="normal", positive=positive, negative=negative, latent_image=latent, denoise=1.0, ) decoded = VAEDecode(samples=sampled, vae=vae) # Export to ComfyUI JSON format workflow = decoded.source_node.build_json_workflow() # Optionally: print or send to a ComfyUI server import json print(json.dumps(workflow, indent=2))

The generator downloads the node schema, sanitises type names, and rewrites nodes.py with updated classes and helper functions.
The just recipe just run wraps the generation command if you prefer using just.

Thanks to ComfyScript for inspiration and prior art.

Read Entire Article