Nodeloom

aka ComfyUI Lite

ComfyUI's node editor for models behind APIs. Build workflows that chain text, image, video and voice models (OpenAI, Google Gemini, OpenRouter, fal.ai, Replicate, BytePlus ModelArk and anything that speaks the OpenAI API) on the canvas you know from ComfyUI, without local inference: no torch, no CUDA, no model downloads, no GPU. It installs in seconds and runs on a laptop or a small server.

_{An idea → Gemini writes the prompt → Nano Banana draws it → Seedance animates it. This is the second run of the graph: every result is reused, nothing is paid twice.}

Think of it as "ComfyUI Lite". It is an independent project, not affiliated with Comfy Org. The editor in the browser is the official ComfyUI frontend (ComfyUI_frontend, the comfyui-frontend-package wheel, unchanged); the backend is a new, small server that speaks the part of ComfyUI's protocol the frontend uses.

Why

ComfyUI is a great editor, but it comes with a local inference stack. If all you need is to wire up hosted models (an LLM writes the prompt, an image model draws it, a video model animates it) you still install gigabytes of torch, and its API nodes only work through Comfy's own cloud account. This project keeps the editor and replaces everything behind it:

• Your providers, your keys. Configure OpenAI, Google's Gemini API, OpenRouter, fal.ai, Replicate, ModelArk, a local Ollama or LM Studio, or your own gateway. Model lists come from each provider's GET /models; every fal.ai or Replicate model you list becomes a node of its own, built from the model's API schema (Flux, Kling, Veo, Wan, Hailuo, ElevenLabs, Whisper and thousands more).
• Parallel by default. Independent branches of a graph run at the same time (two videos from two frames generate together).
• Paid results are never paid twice. Every generation is cached by its inputs: run the graph again and only what changed is regenerated. Change the prompt of the last node and the upstream picture is reused for free.
• Robust queue. Runs wait in a database (SQLite, or PostgreSQL for several processes). A restart in the middle of a 5-minute video generation does not lose it: the run resumes and follows the same upstream task.
• Costs. After a run, the editor tells you what it cost, per node (from the provider's usage and the model's price).
• The Builder. A side panel where you describe what you want ("a product shot on white, then a 5-second video of it") and an LLM builds the graph for you. Nothing runs until you press Run.
• English and Russian UI, including the nodes.

Quick start

Requires Python 3.12+ and, for the video tools, ffmpeg on the PATH.

pip install nodeloom                # or, without installing: uvx nodeloom --open
export OPENAI_API_KEY=sk-...        # and/or GEMINI_API_KEY, OPENROUTER_API_KEY, FAL_KEY, REPLICATE_API_TOKEN, ARK_API_KEY
nodeloom --open                     # http://127.0.0.1:8188

The latest code: pip install git+https://github.com/ShacoTustra/nodeloom.

Or with Docker:

git clone https://github.com/ShacoTustra/nodeloom && cd nodeloom
docker build -t nodeloom .
docker run -p 8188:8188 -e OPENAI_API_KEY=sk-... -v nodeloom:/data nodeloom

The database and your files live in your user data directory (~/.local/share/nodeloom on Linux, ~/Library/Application Support/nodeloom on macOS, %LOCALAPPDATA%\nodeloom on Windows), or wherever --data-dir / DATA_DIR points.

Providers

The common ones need only an environment variable:

Anything else goes into config.toml (in the working directory or the data directory, or --config path). See config.example.toml:

[providers.local]
type = "openai"                       # the OpenAI API: Ollama, LM Studio, vLLM, Groq, DeepSeek, Together...
base_url = "http://localhost:11434/v1"
models = ["llama3.2", "qwen2.5vl"]    # optional: otherwise GET /models

[providers.local.extra_models."qwen2.5vl"]
inputs = ["image"]                    # what a chat model takes besides text

[providers.openai.extra_models."gpt-image-1"]
price = { input = 5.0, image_input = 10.0, output = 40.0 }   # US$ per million tokens, for the cost shown

[providers.fal]
type = "fal"
api_key_env = "FAL_KEY"
[providers.fal.endpoints."fal-ai/flux/dev"]                  # any model of fal.ai/models, by its endpoint id
price = { request = 0.025 }
[providers.fal.endpoints."fal-ai/kling-video/v2.1/master/image-to-video"]
title = "Kling 2.1 Master"

[providers.replicate]
type = "replicate"
api_key_env = "REPLICATE_API_TOKEN"
[providers.replicate.endpoints."black-forest-labs/flux-schnell"]   # any model of replicate.com, owner/name
price = { image = 0.003 }
[providers.replicate.endpoints."openai/whisper"]                   # a community model: its latest version,
price = { second = 0.000975 }                                       # or "owner/name:<version>" to keep one

[providers.google.extra_models."gemini-3.1-flash-image"]
price = { input = 1.5, output = 30.0 }   # Google prices pictures and video by their output tokens

A gateway that passes Google's own API on (generateContent) works as a google provider with its base_url; Gemini video needs Google's Interactions API, which gateways may not have. A router with OpenRouter's video API (POST /videos, GET /videos/models) adds its video models to the Video node.

In the nodes a model is <provider>/<model>, so the same graph can mix providers. A node whose models no provider has is not offered (and a template that needs it is not listed).

A fal.ai endpoint or a Replicate model becomes a node under Generate / fal / or Generate / Replicate / : its settings are widgets (the common ones in front, the rest under "Show advanced inputs"), its image, video and audio inputs are links, its results are outputs. Its schema is fetched once and kept in the data directory. Replicate's schemas say of a file output only that it is a link: the node guesses whether it is a picture, a video or a sound from the model's settings and description, and kind = "video" (or image, audio) on the model sets it when the guess is wrong.

Neither fal.ai nor Replicate reports a price in its answers, and Google reports tokens only: give the models you use a price to see what a run costs (request per call, image per picture, second per second of run time, input / output per million tokens).

Nodes

Images, video and sound travel between nodes as files; a generating node shows its result on the canvas, and every result lands in the editor's assets panel.

Security

There is one local user and no sign-in, as in ComfyUI. The server listens on 127.0.0.1 by default and accepts only requests whose Host is a local name (ALLOWED_HOSTS), so a web page cannot reach it through DNS rebinding; requests that change anything must come from the editor's own pages. To use it over a network, put it behind a reverse proxy with authentication, start it with --host 0.0.0.0 and add your host name to ALLOWED_HOSTS. Your provider keys stay on the server.

Settings

Development

uv sync
uv run pytest          # providers are mocked; the video tests need ffmpeg
uv run nodeloom

How it works inside: docs/architecture.md.

Status

Alpha. Working: the editor, the queue, caching and resuming, the nodes above, fal.ai and Replicate models, Google's Gemini API, the Builder, costs, managing uploads. The provider calls are tested against recorded API shapes; a report of a real run that fails is very welcome. Next: several users, ModelArk's own character library.

License

GPL-3.0, like the ComfyUI frontend it serves. "ComfyUI" is a name of Comfy Org, used here only to say what this works with.