ERPK Collection
A monorepo for ERPK's custom ComfyUI nodes, extending ComfyUI's functionality through integrations with various AI services and APIs.
Nodes (109)
Where the model and key live
Actually chat with Claude inside ComfyUI, and keep the thread
Tokens, history, state
Turn 'a cat' into a shootable scene
Get valid JSON out of Claude, every time
The plainest way to put an LLM in your graph
Know what a call costs before you make it
Hand Claude a schema without writing one
See exactly what Claude cost you, down to the cent
Up to 20 images, one sharp pair of eyes
Edit an image with a prompt, no mask painting
Cloud t2i that speaks text well
Glue your text nodes together
One node to actually see what your graph made
One seed to rule them all, with clamping
The key plumber for every Gemini node
Talk to Gemini and get text you can actually wire into your graph
Object detection that names what it finds
Edit an image by describing the change — no inpainting mask math
Nano Banana power, right in ComfyUI
Get your character into a shot without training a LoRA
The API video node that doesn't poll for half an hour
The content-filter dial for your calls
Set the rules once, apply to everything
The workhorse LLM node of the pack
Ask questions about an image, get answers back
Hand out an API client
A real multi-turn conversation inside your graph
Sentence-driven edits on up to three source images
Grok images without leaving ComfyUI — and n up to 4 costs the same
Up to 3 images become a motion clip
Drop an LLM into the middle of your ComfyUI graph
A 15-second Grok video from ComfyUI, with zero VRAM spent
Edit an existing video by describing the change
Keep the clip going past its natural ending
A portrait-focused Qwen render, same $0.02 price
Lock a character across scenes before you animate
Start frame, end frame, and sound
Turn a prompt into a billed clip, no GPU required
The fast, cheap way to animate a frame
The Pro-only budget path from prompt to clip
Sound is on by default, and that doubles the bill
The budget Kling tier with optional audio
Animate a still into a scene
The flagship Kling node, in your graph
The sibling with a real seed and longer clips
The one LTX node with a real seed
Short clips with sound, billed through WaveSpeed
6/8/10 seconds, sound on, billed per clip
MiniMax H3 Image Edit
Animate a frame, or interpolate between two
The MiniMax H3 LoRA Stack
Character lock, with a bracket-tag catch
MiniMax H3 Text-to-Image as a node
Picture and stereo audio in one pass
MiniMax H3 Video Edit
MiniMax H3 Video Extend
The one node every ERPK OpenAI workflow starts with
Chat with any OpenAI-compatible model and feed the answer back into your graph
Gpt-image-2 with mask inpainting and continuity
GPT-Image-2 with transparent backgrounds and a revised prompt
Let a reasoning model do the prompt-drawing
Set the model's personality once, apply it everywhere
LLM calls where you actually need reproducibility
Make GPT-5.5 look at your images and tell you what it sees
The MiniCPM pack's ghost node (and what replaced it)
The next-gen Qwen editor with Standard and Pro tiers
The current-gen Qwen, Standard or Pro
Rent the newest Qwen editor for three cents a shot
Three cents, fifteen aspect ratios, no VRAM
Edit an image with Qwen — and steer it with up to three LoRAs
Qwen-Image editing in ComfyUI — but it's an API call, not a local model
Cloud editing with your own LoRAs, up to three at once
The 2511 model for multi-person consistency
Decompose one image into editable RGBA layers
Generate with a trained style, paid per image
Six reference images and one instruction, $0.07 per pass
The premium 20B tier when a $0.02 render isn't enough
96 camera poses for 2.5 cents, no LoRA, no 20B download
Orbit your subject with three sliders
Bilingual prompting, five-cent images, zero VRAM
The ERPK Regional Prompt Builder
Pull one region's mask out of the builder's batch
Animate a frame, optionally end on another
Four model tiers, native audio, up to four references
4k, native audio, and a price tag to match
The 4k T2V that takes reference material too
Rewrite a clip by prompt, chained off any video node
Keep the clip going past its last frame
Cloud image editing that's actually good at text
One edit prompt, 1 to 15 coherent results
The text-and-typography specialist, up to 4K
One prompt, a batch of consistent images, 15 max
The workhorse ByteDance edit node
Edit up to 10 images, emit a consistent batch
ByteDance's flagship image model, from 320px up to 4K
A consistent batch of images from one prompt
High-resolution cloud editing with a 1440px floor
Edit up to 10 images into one consistent set
ByteDance's image model, billed by the image
A consistent image batch in one paid call
Ten reference images, one prompt, and a batch-vs-URL trap
Seedream V5.0 Pro Layers slices your image into objects — at $0.77 a pop
Pick a ratio, pick a tier, pay by the tier
The bridge your local images have to cross
The API-only sequel to the last open Wan
Alibaba's video model, no weights required
Extending a clip you already made
The boring node you can usually skip
ComfyUI Custom Nodes - ERPK Collection
A monorepo for ERPK's custom ComfyUI nodes, extending ComfyUI's functionality through integrations with various AI services and APIs.
Current Version: 2026.10.6 (CalVer)
Repository Structure
ComfyUI-Custom-Nodes/
├── wavespeed/ # WaveSpeed AI integration
│ ├── README.md # Package documentation
│ ├── nodes.py # Core nodes
│ ├── seedream_v4*.py # Seedream V4 nodes (4 variants)
│ ├── seedream_v4_5*.py # Seedream V4.5 nodes (4 variants)
│ ├── seedream_v5_5*.py # Seedream V5.5 nodes (3 variants)
│ ├── seedream_v5_flash*.py # Seedream V5.0 Flash nodes (3 variants)
│ ├── flux_2_pro*.py # FLUX.2 Pro nodes (2 variants)
│ ├── qwen_image_*.py # Qwen Image nodes
│ └── wavespeed_api/ # API integration layer
├── claude/ # Claude API integration
│ ├── README.md # Package documentation
│ ├── nodes.py # Core nodes
│ ├── prompt_enhancer.py # Prompt enhancement node
│ ├── vision_analysis.py # Image analysis node
│ └── claude_api/ # API integration layer
├── gemini/ # Google Gemini API integration
│ ├── README.md # Package documentation
│ ├── nodes.py # Text, chat, vision, detect and image nodes
│ ├── omni_nodes.py # Omni Flash video nodes
│ └── gemini_api/ # API integration layer
├── openai/ # OpenAI API integration
│ ├── README.md # Package documentation
│ ├── nodes.py # Core nodes (Config, Text, Vision, Chat)
│ ├── image_nodes.py # Image generation, editing and SVG nodes
│ └── openai_api/ # API integration layer
├── grok/ # xAI Grok API integration
│ ├── README.md # Package documentation
│ ├── nodes.py # Config, Text, Chat, Image nodes
│ ├── video_nodes.py # Text-to-Video, Ref-to-Video, Edit, Extend
│ └── grok_api/ # API integration layer
├── byteplus/ # BytePlus ModelArk (Seedance, Seedream) integration
│ ├── README.md # Package documentation
│ ├── nodes.py # API Client, Seedream Text to Image and Image Edit
│ ├── video_nodes.py # Seedance Text, Image and Reference to Video, Task Result
│ └── byteplus_api/ # HTTP client, per-model capability tables, media
├── utils/ # String and general utilities
│ ├── __init__.py # Module exports
│ ├── concat_strings.py # String concatenation node
│ ├── image_grid.py # Image grid inputs and file loading, shared by grid nodes
│ └── image_grid_loader.py # Image Grid Loader node
├── settings.py # ComfyUI settings reader for API keys
├── shared_workflows.py # Shared workflows CRUD for multi-user
├── shared_workflows/ # Legacy storage fallback (gitignored)
└── web/ # Frontend extensions
├── erpk_settings.js # API key settings in ComfyUI Settings UI
├── shared_workflows.js # Browse/save/delete shared workflows UI
├── node_migration.js # V1→V3 node type rewriting for old workflows
├── aspect_ratio.js # Aspect ratio display in node titles
├── concat_strings.js # Dynamic UI for concat strings node
├── image_grid.js # Registers the drop grid on the grid nodes
└── image_grid/ # Drop grid widget and its node mount
Available Node Packages
ERPK/WaveSpeedAI
Custom nodes for WaveSpeed AI's image generation and editing APIs.
Category in ComfyUI: ERPK/WaveSpeedAI
ByteDance Seedream V4 Models
- Seedream V4 - Text-to-image generation with configurable dimensions (320-4096px)
- Seedream V4 Sequential - Multi-image generation with cross-image consistency (1-15 images, $0.027/image)
- Seedream V4 Edit - AI-powered image editing with text prompts (up to 10 reference images)
- Seedream V4 Edit Sequential - Multi-image editing with coherent results (1-15 images, $0.027/image)
ByteDance Seedream V4.5 Models
- Seedream V4.5 - Enhanced typography and text rendering for posters, logos, UI (1024-4096px)
- Seedream V4.5 Sequential - Multi-image generation with typography support (1-15 images, $0.027/image)
- Seedream V4.5 Edit - Image editing with enhanced text rendering (up to 10 reference images)
- Seedream V4.5 Edit Sequential - Multi-image editing with typography (1-15 images, $0.027/image)
ByteDance Seedream V5.5 Models
- Seedream V5.5 - Text-to-image with a transparent background option and an alpha
maskoutput ($0.02/image at 1k or 1.5k, $0.035 at 2k) - Seedream V5.5 Edit - Image editing from up to 10 reference images, same outputs (as above, plus $0.008 per input image)
- Seedream V5.5 Layers - Splits an image into a clean base plus one layer and mask per object (
max_imagesx the per-image price, $0.20 at the default of 10 at 1k)
ByteDance Seedream V5.0 Flash Models
- Seedream V5.0 Flash - Text-to-image at a flat price for every resolution tier ($0.027/image)
- Seedream V5.0 Flash Edit - Image editing from up to 10 reference images ($0.027/image, however many images go in)
- Seedream V5.0 Flash Layers - Splits an image into a clean base plus up to 16 layers with masks ($0.459/run)
FLUX.2 Pro Models
- FLUX.2 Pro - Text-to-image at 256-2048px with a seed sent to the API ($0.03/image)
- FLUX.2 Pro Edit - Image editing from up to 3 reference images, with an image grid and an optional output size ($0.06/image)
Every Seedream edit node, Sequential included, also has an image grid. With nothing on images, the node sends up to 10 grid files at their own size. It reads the URL text only when the grid is empty.
MiniMax H3 Image Models
- MiniMax H3 Text-to-Image - 1k/2k images in 15 aspect ratios ($0.02-$0.06/image)
- MiniMax H3 Image Edit - Identity-preserving re-render from 1-9 reference images cited as
<Picture N>($0.03-$0.09/image) - MiniMax H3 LoRA Stack - Config node holding up to 3 LoRA weights; connect to any H3 node to call its
-loratwin
MiniMax H3 Image Edit also has an image grid. With nothing on images, it sends up to 9 grid files at their own size. It reads the URL text only when the grid is empty.
Qwen Image Models
- Qwen Image Text-to-Image - Bilingual text-to-image generation (Chinese/English, max 1536×1536, $0.02/image)
- Qwen Image Edit - Single image editing with bilingual prompts (256-1536px, $0.02/image)
- Qwen Image Edit Plus - Advanced editing with up to 3 reference images ($0.02/image)
- Qwen Image 2.0 Text-to-Image - Next-gen Qwen with Standard and Pro quality tiers
- Qwen Image 2.0 Edit - Next-gen Qwen editing with up to 3 reference images, Standard and Pro tiers
- Qwen Image 3.0 Text-to-Image - Qwen Image 3.0 and 3.0 Pro, 1k or 2k at 15 aspect ratios ($0.03-$0.075/image)
- Qwen Image 3.0 Edit - Qwen Image 3.0 and 3.0 Pro editing with up to 3 reference images, 1k or 2k
- Qwen Image Multiple Angles (96 poses) - Re-renders a subject from one of 8 azimuths, 4 elevations and 3 framing distances ($0.025/image)
Every Qwen edit node (Edit, Edit LoRA, Edit Plus, Edit Plus LoRA, Max Edit, 2.0 Edit, 3.0 Edit, Multiple Angles, Multiple Angles (96 poses), Layered) takes a ComfyUI IMAGE directly, sent inline as base64, so you don't need WaveSpeed Upload Image. The multi-image nodes take an IMAGE batch on reference_images; the single-image nodes take one image on input_image. Max Edit, 2.0 Edit, 3.0 Edit, Edit Plus, Edit Plus LoRA and Multiple Angles also have an image grid. With nothing on reference_images, they send the grid files at their own size. They read the URL text only when the grid is empty. The other nodes read the URL text when their image socket is empty. To build an IMAGE batch from files on disk for those, use the Image Grid Loader node.
Video Generation
Video nodes return a STRING URL that works directly with the Preview Anything node for preview/download. Polling runs at a 10s interval, timing out after 15 minutes on most video nodes, 20 on MiniMax H3 Reference-to-Video and 30 on the Seedance 2.5 family and MiniMax H3 Video Edit, which run longest.
- Bytedance Seedance 2.0 Text-to-Video / Image-to-Video - Native audio-visual generation across four model variants (Seedance 2.0, Turbo, Fast, Fast Turbo). T2V supports reference images/videos/audios; I2V supports optional end frame. Turbo variants are 720p/1080p only.
- Bytedance Seedance 2.5 Text-to-Video / Image-to-Video - Longer clips (4-30s) and 4k output. T2V offers the Seedance 2.5 and Turbo tiers and supports reference images/videos/audios; I2V adds a Spicy tier and an optional ending frame
- Bytedance Seedance 2.5 Video Edit / Video Extend - Rewrite an existing clip from a prompt, or continue one past its final frame. Both take the source as a video URL, so they chain off any video node's output
- MiniMax H3 Text-to-Video / Image-to-Video / Reference-to-Video - MiniMax's hosted H3 edition: picture and native stereo audio in one pass at 768p or 2k, 4-15s. Audio is steered by an
Audio:line in the prompt rather than a toggle. Reference-to-Video takes up to 9 images, 3 videos and 3 audios, each cited in the prompt by bracket tag, supplied as ComfyUI IMAGE / VIDEO / AUDIO inputs or as URLs - MiniMax H3 Video Edit / Video Extend (open-weights edition) - Rewrite a clip's lighting, style or elements, or append a new segment after its last frame. Both take the source as a video URL, so they chain off any video node's output
- MiniMax H3 Text-to-Image / Image Edit / LoRA Stack - 1k/2k image generation and identity-preserving edits from up to 9 references. The LoRA Stack config node feeds up to 3 LoRA weights into any H3 node, which then calls the endpoint's
-loratwin - Alibaba WAN 2.7 Text-to-Video / Image-to-Video / Video Extend - WAN 2.7 with extension support for continuing existing clips. Image-to-Video takes its start and end frames as ComfyUI IMAGE inputs (
first_frame,last_frame) or as URLs - Kling 3.0 Image-to-Video - Kling v3.0 i2v with Standard and Pro quality tiers
- Kling O3 Text-to-Video / Image-to-Video - Kling's flagship O3 with Standard and Pro tiers
- Lightricks LTX 2 Pro Text-to-Video / Image-to-Video - LTX 2 Pro with enum durations (6/8/10s) and optional audio
- WaveSpeed LTX 2.3 Text-to-Video / Image-to-Video - LTX 2.3 with 480p/720p/1080p resolution, 5-20s duration, and seed control
LTX 2 Pro and LTX 2.3 are separate nodes (not a unified dropdown) because their APIs take different parameters — LTX 2 Pro has generate_audio, LTX 2.3 has resolution/aspect_ratio/seed.
The Kling (2.5 Turbo, 2.6, Elements) and LTX 2 Pro nodes expose a seed widget for cache control. These WaveSpeed endpoints have no API seed, so the value is not sent to the API: leaving it on "randomize" forces a fresh, re-billed generation on every queue, while a fixed value lets ComfyUI serve the cached result and skip the call. LTX 2.3's seed, by contrast, is a true generation seed sent to the API.
Installation & Documentation: See wavespeed/README.md
⚠️ Note: For the official WaveSpeed ComfyUI nodes and documentation, see the official WaveSpeed ComfyUI repository.
ERPK/Claude
Claude API integration for text generation, prompt enhancement, vision analysis, and conversational AI.
Category in ComfyUI: ERPK/Claude
Nodes
- Claude API Client - Initialize Claude API connection with model selection (Sonnet 5.5 by default, Sonnet 5, Opus 5.5, Opus 5, Fable 5.2, Fable 5.1, Fable 5, Haiku 5.5) and configuration. Optional if API key is configured in ComfyUI Settings or config.ini -- Claude nodes can run standalone.
- Claude Prompt Enhancer - Transform simple prompts into detailed descriptions with 51 artistic styles (photorealistic, cinematic, fantasy, cyberpunk, anime, oil painting, watercolor, and more)
- Claude Vision Analysis - Analyze images with Claude's multimodal capabilities (up to 600 images per request). Optional per-node
modeloverride lets you use a different model for vision than the one your client is configured with. - Claude Text Generation - General-purpose text completion and generation
- Claude Conversation - Multi-turn dialogues with context preservation and automatic memory management
- Claude Conversation Info - Display conversation statistics and token usage
- Claude Tool Definition - Build Anthropic tool definitions for structured output (chainable)
- Claude Structured Output - Get JSON matching a tool schema via Anthropic structured outputs
- Claude Token Counter - Count tokens and estimate API costs before making requests
- Claude Usage Stats - Track cumulative token usage and costs across all Claude nodes
Key Benefits:
- 51 artistic styles for prompt enhancement (photorealistic, cinematic, fantasy, anime, oil painting, impressionist, cyberpunk, and more)
- Claude 5-generation models only: 1M context, up to 128K output, adaptive thinking, and
effortandthinking(adaptive or off) widgets on every generation node - Automatic prompt caching (cache reads at 0.1x input on most models)
- Usage Stats priced per answering model, including cache reads and writes
- Streaming support for real-time responses
- Automatic context window management
- Multi-image analysis capabilities
- Cost optimization with token counting
Installation & Documentation: See claude/README.md
ERPK/Gemini
Google Gemini API integration for text generation, vision analysis, multi-turn conversations, image generation, image editing, and Omni Flash video generation.
Category in ComfyUI: ERPK/Gemini
Nodes
- Gemini API Config - Initialize Gemini API connection (API key configuration). Optional if API key is configured in ComfyUI Settings or config.ini -- Gemini nodes can run standalone.
- Gemini Text Generation - General-purpose text generation with model selection (Gemini 4 Hydrogen, 4 Carbon, 4 Argon, 3.1 Pro Preview, 3.8/3.7/3.6/3.5 Flash, 3.5 Flash-Lite)
- Gemini Chat - Multi-turn conversations with automatic context preservation
- Gemini Vision - Analyze images with multimodal capabilities. Vision and Detect take a
media_resolutionwidget;lowspends about a quarter of the image tokens ofhigh - Gemini Detect - Open-vocabulary object detection: name objects (or leave empty to detect everything prominent) and get normalized regions on an
ERPK_REGIONSoutput that wires into the Regional Prompt Builder'sregionsinput - Gemini Image Generation - Generate images from text descriptions (Nano Banana 2.1, 3.1 Flash, 3.1 Flash Lite, 3 Pro; 0.5K to 4K resolution, up to 14 aspect ratios, Google Search grounding on all but 3.1 Flash Lite)
- Gemini Image Edit - Edit and modify images with natural language prompts (up to 14 reference images, same model and resolution options). Connect an IMAGE, or pick or drop up to 14 files onto the grid on the node, and each file goes to the model at its own size
- Gemini System Instruction - Set persistent system-level instructions to guide model behavior
- Gemini Safety Settings - Configure content safety filters (strict/balanced/permissive presets or custom)
Video Generation Nodes
- Gemini Omni Video Generation - Generate video with sound of a chosen length, 3-15s (360p to 4k), from a prompt or start image via Gemini Omni Flash (returns in one call; 1080p/4k arrive through the Files API)
- Gemini Omni Reference to Video - Generate video featuring the subjects in reference images via Gemini Omni Flash
Key Benefits:
- Support for Gemini 4 Hydrogen (default), 4 Carbon, 4 Argon, 3.1 Pro Preview, 3.8, 3.7, 3.6 and 3.5 Flash, and 3.5 Flash-Lite
- Thinking level control across all models (none/minimal/low/medium/high/xhigh).
nonekeeps the model's default thinking; the node raisesminimaltolowand lowersxhightohighon models that reject them (only Gemini 4 Hydrogen takesxhigh) - Omni Flash video generation from text, a start image or reference images
- Each node selects its own model for maximum flexibility
- State-of-the-art reasoning with Gemini 3.1 Pro Preview
- Image generation with Nano Banana 2.1 (recommended), 3.1 Flash, 3.1 Flash Lite, and 3 Pro (0.5K to 4K resolution, model-dependent)
- Image editing with natural language instructions (up to 14 reference images)
- Simple, straightforward API integration
- Vision capabilities with batch image support
- Flexible safety controls
- Native multi-turn conversation support
Installation & Documentation: See gemini/README.md
ERPK/OpenAI
OpenAI API integration for text generation, vision analysis, multi-turn conversations, image generation, image editing, and SVG generation.
Category in ComfyUI: ERPK/OpenAI
Nodes
- OpenAI API Config - Initialize OpenAI API connection (API key configuration). Optional if API key is configured in ComfyUI Settings or config.ini -- OpenAI nodes can run standalone.
- OpenAI Text Generation - General-purpose text generation with model selection (GPT-6.2 Sol default, GPT-6.2 Astra, GPT-6.1 Sol/Terra, GPT-6 Astra/Sol/Luna, GPT-5.6 Sol/Terra/Luna, chat-latest) and optional
reasoning_effortandverbositycontrols - OpenAI Chat - Multi-turn conversations with automatic context preservation, plus reasoning_effort and verbosity control on the gpt-6.x and gpt-5.6 models
- OpenAI Vision - Analyze images with any OpenAI text model (GPT-6.2 Sol default)
- OpenAI System Instruction - Set persistent system-level instructions to guide model behavior
- OpenAI Image Generation - Generate images with GPT-Image-2.6 Flare (default), GPT-Image-2.6 Sunburst, GPT-Image-2.5 Flare / Sunburst and GPT-Image-2, with png / jpeg / webp output, plus animated gif on 2.6 (every frame lands in the IMAGE batch) and a
complexitycontrol on 2.6 - OpenAI Image Generation (Responses) - Generate images via the Responses API with a mainline reasoning model (gpt-6.2-sol default) driving the
image_generationhosted tool (GPT-Image-2.6 Flare by default). Adds optional web search grounding and prompt revision. - OpenAI Image Edit - Edit and inpaint images with natural language prompts (GPT-Image-2.6 Sunburst default). Connect an IMAGE, or pick or drop up to 16 files onto the grid on the node, and each file goes to the API at its own size. A markup layer (arrows, circles or text drawn on a transparent PNG) shows the model where and what to edit
- OpenAI SVG Generation - Generate a vector image as SVG markup with GPT-Image-2.6, which hands the drawing to a text model (gpt-6.2-sol default); the output is one STRING
Key Benefits:
- Support for GPT-6.2 Sol (default: 1.05M context, 128K output, $2/$10 per MTok) and GPT-6.2 Astra (same context and output, $8/$40 per MTok), GPT-6.1 Sol ($2/$10 per MTok) and GPT-6.1 Terra ($1/$5 per MTok), the GPT-6 Astra/Sol/Luna and GPT-5.6 Sol/Terra/Luna tiers, and chat-latest. The GPT-5.5 and older text models were removed on 2026-10-01
reasoning_effortparameter (default/none/minimal/low/medium/high/xhigh) for the reasoning models, matching ComfyUI's own OpenAI node;defaultleaves the effort to the model,noneturns reasoning off on gpt-6.1-terra, gpt-6-sol, gpt-6-luna and the gpt-5.6 tiers and acts likedefaulton the gpt-6.2 tiers, gpt-6.1-sol and gpt-6-astra,minimalis sent aslow, and chat-latest drops it. The Responses image node also offersmaxand aproreasoning modeverbosityparameter (default/low/medium/high) — shapes how chatty the response is independently ofmax_tokens; not sent to chat-latest, which ignores it- Every current model rejects temperature, top_p and stop while reasoning; temperature and top_p are sent only with
reasoning_effortnone on a model that accepts it, and stop is never sent - Image generation with GPT-Image-2.6 Flare / Sunburst (complexity control, gif and SVG output; Flare is the generation default, Sunburst the edit default), GPT-Image-2.5 Flare / Sunburst (both tiers of 2.6 and 2.5 add xhigh/max quality and transparent background) and GPT-Image-2 (4K output, multilingual text). GPT-Image-1.5, 1 and 1-Mini were removed on 2026-10-01
- Image editing with an optional mask for inpainting, or a markup layer of drawn marks
- Multi-turn conversation with session management
- Automatic retry with exponential backoff
- JSON mode and strict json_schema Structured Outputs on Text Generation, Chat and Vision
Installation & Documentation: See openai/README.md
ERPK/Grok
xAI Grok integration — text, multi-turn chat, image generation/editing (up to 5 source images), and full video coverage (text-to-video, reference-to-video, edit, extend).
Category in ComfyUI: ERPK/Grok and ERPK/Grok/Video
SDK requirement: xai-sdk>=1.20.0
Text & Chat Nodes
- Grok API Client - Initialize the xAI client (optional for downstream nodes)
- Grok Text Generation - One-shot text completion via
grok-4.7/grok-4.20-0309-non-reasoning/grok-4.20-multi-agent-0309/grok-build-0.1, withreasoning_effortfor grok-4.7, agent count (4 or 16) for grok-4.20-multi-agent-0309, an optionalsystem_prompt, and aseedthat reaches xAI from 1 up - Grok Chat - Multi-turn conversation threaded via
GROK_CHAT_SESSION, with an optionalsystem_promptthat the session carries to later turns and the sameseed - A Grok Text Generation or Chat reply cut off at
max_tokensraises an error instead of returning partial text
Image Nodes
- Grok Image Generation - Text-to-image via
grok-imagine-image-2.0, 13 aspect ratios plus auto, 1k/2k resolution, low/medium quality, n=1..10 batched output - Grok Image Edit - Single or multi-image editing (up to 5 source images per xAI's cap; the node rejects a larger batch). Connect an IMAGE, or pick or drop up to 5 files onto the grid on the node, and each file goes to xAI as its own source image in grid order
Video Nodes
- Grok Text to Video - Text-to-video via grok-imagine-video-1.5-lite (default), grok-imagine-video-1.5 or grok-imagine-video, 1-15s, 7 aspect ratios, 480p/720p/1080p (1080p on 1.5 and 1.5-lite)
- Grok Reference to Video - Up to 14 (grok-imagine-video-1.5, default, $0.08/s) or 7 (grok-imagine-video, $0.05/s) reference images guide generation;
<IMAGE_0>,<IMAGE_1>, ... token addressing - Grok Video Edit - Edit an existing video (URL or data URI) with a text prompt (output capped at 8.7 s and 720p)
- Grok Video Extend - Append 2-10 seconds of new content to an existing video
Installation & Documentation: See grok/README.md
ERPK/BytePlus
Seedance video and Seedream image generation called directly on BytePlus ModelArk with your own API key, over plain HTTPS. These nodes do not go through WaveSpeed, so the bill and the model activation are on your BytePlus account.
Category in ComfyUI: ERPK/BytePlus and ERPK/BytePlus/Video
- BytePlus API Client - Optional config node holding the ModelArk key; without it the nodes read ComfyUI Settings or
byteplus/config.ini - BytePlus Seedance Text to Video / Image to Video / Reference to Video - All seven current Seedance models: 2.5, 2.5 flash, 2.0, 2.0 fast, 2.0 mini (default), 1.0 pro and 1.0 pro fast. Image to Video takes a start frame and an optional end frame; Reference to Video takes reference images plus reference video and audio URLs. Every documented option is a widget except
callback_url,streamand the region (see the provider README): resolution up to 4k, ratio incl. adaptive, duration, 1.0 frames, seed, camera_fixed, audio, watermark, last frame, mov output, flex tier, priority, 2.5 drafts and draft renders, task expiry and a custom endpoint ID. Each outputs the video URL, the last-frame URL and the task ID - BytePlus Seedance Task Result - Reads a Seedance task's video URL, last-frame URL and status by task ID, optionally waiting for it to finish. It only reads, so it bills nothing; use it to collect a video from a cancelled or lost run
- BytePlus Seedream Text to Image / Image Edit - Seedream 5.5 (default), 5.0 pro, 5.0 flash, 5.0 lite, 4.5 and 4.0: size presets or an exact size, sequential batches of up to 15, PNG or JPEG, transparent backgrounds, prompt optimization mode and watermark. Image Edit adds up to 14 references, a mask for reference alpha, and layer decomposition. Both output an IMAGE batch plus its alpha
A combination the selected model rejects stops before any call, with a message naming the model and the value. A seed reproduces a result on Seedance 1.0 only, so elsewhere the seed widget is cache control only.
Installation & Documentation: See byteplus/README.md
Background Removal (Removed)
Background removal nodes have been removed from this package. For background removal in ComfyUI, use ComfyUI-RMBG instead.
ERPK/utils
String manipulation and general utility nodes.
Category in ComfyUI: ERPK/utils
Nodes
- Concatenate Strings - Combine multiple text inputs with configurable delimiters. Supports up to 10 connectable inputs - drag STRING outputs from other nodes or enter text directly.
- Image Grid Loader - Pick or drop up to 14 image files onto a grid on the node and get them back as one IMAGE batch in grid order, ready for any multi-image edit node. Cell 3 is image 3 in the batch.
- Seed - Generate a seed value with optional min/max range clamping. Connect the output to any node's seed input to share a single seed.
- Preview Anything - Preview any value: text, markdown, image/video/audio/gif URLs, IMAGE tensors, and AUDIO dicts. Includes a Download button that saves the rendered content to your computer.
- Regional Prompt Builder - Draw regions on a canvas and emit a layout-aware prompt for any image generation model: each region becomes a verbal placement plus box_2d coordinates (works with the Gemini, OpenAI, and Grok image nodes). Also outputs pixel-space
BOUNDING_BOXregions compatible with core SAM3 Detect, Draw BBoxes, and Crop By BBoxes nodes.
Regional Prompt Builder features:
- Canvas editor on the node: drag to draw, move, and resize regions; shift-click or a shift-drag marquee selects multiple regions and dragging any selected region moves the whole selection; Ctrl/Cmd-drag force-draws over existing boxes; Alt-click cycles through overlapping regions; Delete removes the selection; Ctrl/Cmd+C/V/D copy, paste, and duplicate; H hides the boxes for a clear view of the frame; F (or the ⤢ button) expands the editor to fill the window (Esc exits). An inspector row under the canvas edits the selected region's description, kind (object or rendered text), and literal text live; double-click a region to jump to its description field
- Exact geometry: right-clicking a region selects it and the region list shows X/Y/W/H fields for it in frame pixels, applied live (right-clicking empty canvas opens just the list); arrow keys nudge the selection by 1px (10px with Shift) and Alt+arrows resize
- Right-click the canvas for a region list (top = front): click a row to select, drag rows to reorder depth, duplicate or delete per row
- Regions keep identity colors; text regions preview their literal text in-frame
- Depth control: regions layer back to front (number 1 is backmost). Reorder with
[/]or the inspector's depth buttons; the prompt tells the model that later elements appear in front where regions overlap - Placement anchoring: hand-drawn additions and model-applied moves are described relative to the nearest grounded objects ("between the hippo and the man", or "in front of the man" when overlapping), because edit models follow landmarks far better than box_2d coordinates - which they treat as a weak hint and override for physical plausibility (a hat boxed in mid-air lands on the ground). Only objects already in the image (scanned or node-composited) anchor a placement; for an object with no support under its box, draw the box on a real surface
- Deterministic object insertion: wire an object's image into a region's
refsocket and set the region to Node - with an image connected, the reference is composited into the scene at the region's box (background auto-keyed, subject fit and centered) so the box is authoritative for placement and the edit model only blends it; set Model (or run with no image) to instead have the edit model reproduce and place the reference itself (coordinate-blind). The only reliable way to place an added object pixel-exact - Status strip shows region count, selection, frame dimensions, and aspect ratio
- A
?button in the status strip toggles a shortcuts cheat sheet over the canvas (Escape or an outside click dismisses it) - Clear button in the status strip empties all regions (two-step confirm, no blocking dialog)
- Optional grid with a typed cell size in frame pixels (8-1024, default 64 - latent-friendly), a color picker plus opacity control, and snap-to-grid; drawing, moving, and resizing quantize to pixel multiples, and the preferences save with the workflow
- Optional reference image input: the connected image displays under the regions (LoadImage shows immediately) and passes through unchanged on an
imageoutput, so the builder sits inline in image-edit chains. When the image's aspect differs from the frame, a⚠ matchchip in the status strip sets width/height to the image with one click - Dynamic descriptions: optional
desc_1-desc_10STRING sockets override the matching region's description at execute time (numbered as on the canvas), so any text node can drive a region. Sockets stay hidden until exposed with the inspector's plug button (or already wired); pressing the button again hides an empty socket or disconnects a wired one (unsetting it from the node face); socket labels carry the region's text, and the inspector locks a wired region's description field - Per-region reference images: optional
ref_1-ref_10IMAGE sockets attach a reference to the matching region (inspector ▣ button, same flow as descriptions; 10 matches Gemini's high-fidelity object budget). Wired refs flow out on theimage_refsoutput in region order; connect it to Gemini Image Edit'simage_refsinput and each region's prompt line cites its image number, with the edited image always image 1. Ref-wired regions show a ▣ chip and a corner thumbnail - Wired regions: an optional
regionsinput accepts detected regions (e.g. from Gemini Detect) and appends them after the canvas-drawn regions at execute time - the canvas stays untouched,desc_N/ref_Nkeep binding canvas regions only, and appended regions render in front where they overlap - Undo/redo for everything region-shaped (Ctrl/Cmd+Z / Shift+Z): draws, moves, deletes, typing, depth changes, scans, and Clear all each revert as sensible single steps, 50 deep
- Object scan: with an image connected, a ✦ button in the canvas corner has Gemini detect the scene's objects - each becomes a real editable region with a layer name (the label) and a per-region prompt (a one-sentence caption that feeds generation), depth pre-ranked back-to-front, while SAM computes segmentation masks locally (weights download once). Masks keep their largest blob, fill holes, and feather their edges; small objects are re-segmented from a cropped close-up for full-resolution edges. The scan replaces the canvas as a single undo step. A ⚙ options window picks the Gemini detection model and the segmentation backbone (saved with the workflow): the segmenter list is filtered to the SAM families this install can run (SAM / SAM-HQ / SAM 2.1,
↓= downloads on first use) and defaults to SAM ViT-Base - Scan cost readout: each scan estimates its price from the API's reported token usage and shows it in the status strip (last scan plus a running session total) with a toast after the scan finishes; it is a paid-tier estimate (the free tier bills nothing) at the rate in effect on the scan's UTC date, and reads "n/a" for a model with no price on file
- Layer groups: Alt-drop a list row to nest regions; parents collapse, group-hide, and move with their children, and section scans group their detections under the scanned region
- Layer-style region list: right-clicking a region opens a detail view - mask thumbnail, editable name and prompt, exact pixel geometry, per-layer hide/show eye, delete - above the reorderable list, which captions rows by layer name and adds per-row and global hide/show toggles
- A
masksoutput emits a frame-sized MASK batch (one per region, in region order): scanned regions use their segmentation, hand-drawn regions get filled rectangles - ready for inpainting chains; the Region Mask node picks a single region's mask out of the batch by canvas number - Mask-aware selection: hovering the canvas glows the object's mask under the cursor, and clicks pass through the empty corners of scanned regions to whatever is really under the pointer - overlapping objects stay individually clickable
- Object move/scale: drag a scanned object and its masked cut-out follows live with a dashed ghost at the origin; at execute time the move is composited into the image output and the original location is inpainted away (OpenCV), both deterministically - so a moved object never appears twice - and the prompt asks the edit model only to blend the pasted copy
- Move behind another object: a node-composited move lands behind any scanned region (static or itself moved) that sits above it in the depth order and overlaps it - the upper object's pixels end up on top (re-stamped for a static occluder, painted last for a moved one), so the move reads as going behind it; the canvas preview shows this same layering for a Node move, pixel-for-pixel against the deterministic
imageoutput (a Model move previews the intended result - the object stamped at its destination behind the same occluders, with a from/to crosshair pair and a half-opacity origin checkerboard the object shows through, since the node leaves the original in place for the edit model), and the prompt is framed from the occluder's side ("with the man in front of it, the lemur is partly hidden") to stop the edit model redrawing the covered part. Order the occluder above the moved object in the layer list (it must be a scanned region with a mask; a hand-drawn box is a layout guide, not an occluder). Occlusion is deterministic only in Node mode - a Model-mode move gets the same clause but it is best-effort, because the edit model decides depth and tends to render subjects fully visible - Move applied by Node or Model (per-region toggle in the detail panel): Node (default) composites the move into the image pixel-exact and clears the origin, so position is precise and the prompt asks the edit model only to relight, ground with a contact shadow, and match depth of field / grain / ambient color so it doesn't look pasted; Model keeps the object in the image (the node does not clear its origin) and asks the edit model to relocate it from its origin marker (A) to its destination marker (B), erasing the original and rebuilding that spot as background - it stays in context so identity/lighting carry over with no reference crop, but lands only approximately (edit models position by scene semantics and the markers, not by box_2d); this trusts the model to remove the original, so if it leaves a duplicate switch that region to Node for an exact move. To steer Model placement the prompt anchors the destination to the nearest named objects - "between the tiger cub and the man" when they straddle the spot, or "behind the man, in front of the cub" (from the layer order) when it overlaps them, so depth is respected too; for an exact position use Node
- Placement markers (per-region Marker toggle in the right-click detail panel, on by default, needs a connected image): draws a high-contrast solid dot - a filled colored disc over a dark halo with a label letter (A, B, C…) centered inside - for every element the edit model acts on by reading pixels: an addition gets one dot at its spot ("center it on marker A - the magenta dot"), a Model move gets two (one on the object at its origin, one at its target) cited as "move it from marker A to marker B", and a Model cut-out gets one dot on the object to delete ("the object at marker C … remove it"). This is visual prompting (set-of-mark): an edit model ignores box_2d coordinates but reads pixels, so a bold dot it can see is a far stronger placement cue than coordinates it can't. Each marker carries two independent identifiers - a distinct unnatural color (magenta, cyan, lime, …, skipping any color word already in the description) and a unique letter - so markers can't be confused. The dot sits under where the element should land, so a hit hides it and the prompt asks the model to paint it out (a surviving dot signals the model ignored the spot). Node-composited and inserted regions never get one, and a region's Marker can be switched off individually. It raises the odds, not a guarantee - for position-critical work use Node
- Cut out (Shift+Delete): removes the selected region and inpaints its masked area from the surrounding pixels (OpenCV) in the
imageoutput, and the prompt instructs a downstream edit model to rebuild that area as natural background with no new objects - so the object is removed and the gap filled naturally whether you save the image directly or run it through an edit node (output stays RGB); marked on the canvas with the transparency checkerboard behind the mask; Ctrl/Cmd+Z restores it - Removal fill: the
removal_fillwidget chooses how cleared areas (cut-outs and a moved region's old position) are filled -inpaint(default) rebuilds the background with OpenCV, orchromalays a flat key color (thechroma_colorwidget, default chroma green#00B140) so a downstream chroma-key/compositing node can replace it cleanly instead of relying on the inpaint, which can smear on busy scenes; when chroma is active the prompt names the key color so an edit model treats those patches as empty to rebuild - The prompt instructs the model to treat placement areas as invisible guides so boxes are not rendered into the image
Concatenate Strings features:
- 10 connectable text inputs (Text 1 through Text 10)
- Optional labels for each input (Label 1 through Label 10)
- Configurable delimiter with escape sequence support (\n, \t)
- Label placement options (same line or new line)
- Add/Remove buttons to dynamically manage inputs
- Visual separators between configuration and input sections
Image Grid Loader features:
- Pick files with Choose files… or the + Add cell, or drop them onto the grid; it takes PNG, JPG, and WebP
- The grid uploads each file to ComfyUI's
inputfolder through the standard image upload, so the workflow saves file names, not pixels - Numbered cells in send order; drag to reorder, click to see full size, hover × to remove, Clear all with Undo
- A Reference images header with the count collapses the grid to one row; the workflow saves the collapsed state
- Each file shows as a numbered placeholder with a spinner while it uploads, and lands in drop order; PNG thumbnails load through ComfyUI's WebP preview
- Press Tab to focus a cell; the arrow keys move it, Enter shows it full size, and Delete or Backspace removes it
- The grid holds up to 14 files; it ignores dropped files past that and the overlay says how many free slots remain
- The node centers mixed sizes on a canvas of the largest height and width with mid-grey padding, since a batch must share one size
- Calls no API; a missing or out-of-folder file fails when you queue, naming the entry
Preview Anything features:
- Accepts any value via a wildcard input
- Auto-detects type (image/video/audio by URL extension, markdown by syntax, IMAGE tensor, AUDIO dict)
- Dark, distraction-free preview surface
- Collapsible Options panel tucks
display_type,filename, andstrip_metadatabehind a single bar so the node stays clean; click to expand the styled controls - Toolbar appears only when there is content — the Download / Copy buttons and character counter stay hidden until a payload is rendered
display_typedropdown to force a specific renderer- Download button saves content as
.txt,.md, or the original media format - Copy button in the toolbar: copies text/markdown to clipboard, or copies the current image (single-image and gallery-single views) as a PNG blob
- Character counter appears in the toolbar for text and markdown output
- Image dimensions badge (W × H) overlays the bottom-right of single-image previews
- Text and markdown stay at a fixed scrollable size (no auto-growth to canvas-swallowing heights); images, video, and audio still auto-fit to their aspect ratio
- Last rendered content persists across workflow reloads
- IMAGE previews embed the prompt and workflow like ComfyUI's Preview Image, so a downloaded PNG loads back as a workflow. The
strip_metadatatoggle saves them without it, and re-encodes image URL inputs to remove EXIF / ICC / XMP (GPS, camera info, timestamps) before download.
Shared Workflows (Multi-User)
In multi-user ComfyUI (--multi-user), each user's workflows are sandboxed. Shared Workflows provides a common directory where any user can save, browse, load, and delete workflow templates. Also works in single-user mode.
Access: Right-click the canvas > ERPK submenu, or open ERPK Settings to see the inline shared workflows list below your API keys.
- Browse Shared Workflows... - View all shared workflows with name, size, date, and authorship (created by / last edited by). Load any workflow into your canvas or delete it.
- Share Current Workflow... - Save the current canvas workflow to the shared directory. The name field is pre-populated from the current workflow tab.
- Save to "[name]" - Appears after loading or sharing a workflow. Saves directly to the linked shared workflow without opening a dialog.
- Settings panel - The ERPK Settings panel shows an inline shared workflows list with Load, Delete, and Share Current buttons.
Linked save-back: When you load a shared workflow or share one, the workflow name is linked. Subsequent edits can be pushed back to the shared copy via the "Save to" menu item. Other users see the updated version when they browse.
Authorship tracking: Each shared workflow records who created it and who last modified it. On overwrite, the original creator is preserved while the modifier is updated.
Toast notifications: Success/error feedback appears as a brief toast in the top-right corner for all save, load, and delete operations.
Storage: Workflows are saved as JSON files in shared_workflows/ under the ComfyUI base directory (e.g., ~/Documents/ComfyUI/shared_workflows/), alongside input/, output/, and models/. This location persists across plugin upgrades and ComfyUI restarts. Falls back to the extension directory when running outside ComfyUI. Each file uses an envelope format ({meta, workflow}) that stores authorship metadata alongside the workflow graph. No API keys or user-specific data are stored.
Write protection: Every route that writes or deletes a file (save, delete, restore, purge) requires the X-ERPK-Write-Token header carrying the token served by GET /erpk/write_token. The token is generated once per server process. A page from another origin can send a request to the server but cannot read that token, and the routes also refuse a request whose Origin does not match the Host or that the browser marks Sec-Fetch-Site: cross-site. The bundled UI fetches the token on first use. Scripts calling the API directly fetch it the same way.
API Endpoints:
| Endpoint | Method | Description |
|----------|--------|-------------|
| /erpk/write_token | GET | Token required by every write route below |
| /erpk/shared_workflows | GET | List all shared workflows (metadata including authorship) |
| /erpk/shared_workflows/{name} | GET | Get a single workflow by name |
| /erpk/shared_workflows | POST | Save a workflow ({name, workflow}); records user as author. Token required |
| /erpk/shared_workflows/{name} | DELETE | Delete a workflow by name (moves it to trash). Token required |
| /erpk/shared_workflows/trash | GET | List trashed workflows |
| /erpk/shared_workflows/trash/{trash_id}/restore | POST | Restore a trashed workflow. Token required |
| /erpk/shared_workflows/trash/{trash_id} | DELETE | Purge a trashed workflow. Token required |
SaveImage Metadata Toggle
Adds a strip_metadata toggle to ComfyUI's built-in SaveImage node (and PreviewImage, which inherits it). When enabled, both the workflow JSON and prompt data are stripped from saved PNG files, producing clean images with no embedded ComfyUI metadata.
- Per-node control - Each SaveImage node gets its own boolean toggle; no global setting needed
- Off by default - Metadata is preserved unless you explicitly enable stripping
- Prompt metadata also stripped - Both
extra_pnginfo(workflow graph) andprompt(API-format node data) are removed when enabled
Auto-Clear Job History
Automatically removes completed jobs from the history panel after each run. Prevents the UI from slowing down during long sessions with many queued generations.
- Off by default - Enable via Settings > ERPK > General > Auto-Clear Job History
- Delayed cleanup - Waits 5 seconds after job completion so the frontend can fetch results before the history entry is removed
- No restart needed - Toggling the setting takes effect on the next completed job
Installation
Method 1: ComfyUI Manager (Recommended)
- Open ComfyUI Manager
- Click Install Custom Nodes
- Search for
erpk - Find ERPK Custom Nodes and click Install
- Restart ComfyUI
Method 2: ComfyUI Registry
Install directly from the ComfyUI Registry web interface.
Method 3: Manual Installation
-
Navigate to your ComfyUI custom_nodes directory:
cd /path/to/ComfyUI/custom_nodes/ -
Clone this repository as
erpk:git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk -
Install dependencies:
cd erpk pip install -r requirements.txt -
Restart ComfyUI
Post-Installation
-
Configure API keys for the services you want to use. The easiest way is via right-click on the canvas > ERPK Settings, or Settings > ERPK > API Keys in ComfyUI -- keys are stored per-user and never saved in workflows. In the ERPK Settings panel, API-key fields show a masked preview (e.g.
sk-a…kgAA) and a small Edit affordance; click a key to reveal the input, blur to re-mask.API keys are resolved in priority order:
- ComfyUI Settings (recommended) -- right-click canvas > ERPK Settings
- Widget input -- api_key field on client/config nodes
- Config file --
provider/config.ini
Multi-user support: In multi-user ComfyUI installations, each user's API keys are stored separately in their own
comfy.settings.json. The correct user's keys are resolved automatically during workflow execution. A "Current user: [name]" banner appears in the ERPK Settings panel when multi-user mode is active.- WaveSpeed: See wavespeed/README.md
- Claude: See claude/README.md
- Gemini: See gemini/README.md
- OpenAI: See openai/README.md
- Grok: See grok/README.md
- BytePlus ModelArk: See byteplus/README.md
-
Find nodes under their respective categories:
ERPK/WaveSpeedAI,ERPK/Claude,ERPK/Gemini,ERPK/OpenAI,ERPK/Grok,ERPK/BytePlus, andERPK/utils
Backward Compatibility
This package uses the ComfyUI V3 node API. Saved workflows that reference older node IDs (e.g. "WaveSpeed Custom SeedreamV4") are automatically migrated to the current node IDs at load time via ComfyUI's NodeReplace system. No manual workflow editing is required.
ComfyUI API Integration
ComfyUI provides a REST API that allows programmatic workflow creation and execution. This is useful for automation, testing, and integration with external tools.
Available Endpoints
| Endpoint | Method | Description |
|----------|--------|-------------|
| /system_stats | GET | System information (OS, RAM, GPU, versions) |
| /object_info | GET | List all available nodes and their input/output types |
| /prompt | POST | Queue a workflow for execution |
| /queue | GET | View pending and running jobs |
| /history | GET | View execution history and results |
| /history/{prompt_id} | GET | Get results for a specific execution |
Submitting a Workflow via API
curl -X POST http://localhost:8188/prompt \
-H "Content-Type: application/json" \
-d '{
"prompt": {
"1": {
"class_type": "GeminiAPIConfig",
"inputs": {
"api_key": ""
}
},
"2": {
"class_type": "GeminiTextGeneration",
"inputs": {
"client": ["1", 0],
"prompt": "Write a haiku about ComfyUI",
"model": "gemini-3.8-flash",
"temperature": 0.7,
"max_tokens": 256
}
},
"3": {
"class_type": "PreviewAny",
"inputs": {
"source": ["2", 0]
}
}
}
}'
Note: The port may vary (8000 for desktop app, 8188 for standard installation).
Workflow JSON Format
Workflows can be saved as JSON files in your ComfyUI workflows directory. The format includes:
nodes: Array of node definitions withid,type,pos,inputs,outputs, andwidgets_valueslinks: Array of connections in format[link_id, source_node, source_slot, target_node, target_slot, type]last_node_id/last_link_id: Tracking for ID generation
Checking Available Nodes
# List all ERPK nodes
curl -s http://localhost:8188/object_info | jq 'keys' | grep -i -E "(gemini|claude|openai)"
License
MIT License