LiteLLMMessage

Before you can ask a model anything, you need a conversation to send it. That's what LiteLLMMessage is for: it's the "make a chat message" node in the ComfyUI_LiteLLM pack, and it's usually the first node you drop on the canvas when you're building an LLM workflow.

The pack is a bridge between ComfyUI and LiteLLM, the Python library that normalizes ~100 model providers behind one API. So the messages this node builds are provider-agnostic - the same user/assistant/system structure works whether you're calling Claude, GPT, Gemini, or a local Ollama model.

How it works

A chat message is just a dict with two keys: role and content. That's all this node produces. Give it:

• content - the actual text, a big multiline field (defaults to "Hello World!").
• role - a dropdown with three choices: user, assistant, system.

Run it and you get one LLLM_MESSAGES object out, which is the pack's typed way of saying "a list of chat messages." That's the important bit: LiteLLMMessage doesn't call any API, doesn't need a key, costs nothing. It's pure data construction.

The trick that makes it useful is the optional messages input. Plug an existing LLLM_MESSAGES chain into it and the node appends its new message to the end of that chain instead of starting fresh. That's how you build a real conversation:

LiteLLMMessage (system, "You are a helpful editor.")
  → LiteLLMMessage (user, "Tighten this paragraph.")
    → LiteLLMMessage (assistant, "Here's my rewrite.")
      → LiteLLMMessage (user, "Now do the same for the rest.")
        → ShowMessages   (inspect what you've built)

Each node hands its output to the next one's messages input. One nicety: if content is empty, the node returns the incoming messages unchanged - so you can use it as a passthrough while you sketch the workflow.

Wiring it up

The Message(s) output feeds any node in this pack that accepts LLLM_MESSAGES - the completion nodes are the usual destination, and ShowMessages/ShowLastMessage are handy for checking what you've actually built before you spend tokens on it. If you want the text out as a plain string instead, run it through MessagesToText first. And don't hand the output to a plain STRING input on another pack's node - it's a list of dicts, not text.

Installing

LiteLLMMessage ships in ComfyUI_LiteLLM, so you install the whole pack:

• ComfyUI Manager → Custom Nodes Manager → search "LiteLLM" (repo Hopping-Mad-Games/ComfyUI_LiteLLM) → Install → Restart.
• Or manually: cd ComfyUI/custom_nodes && git clone https://github.com/Hopping-Mad-Games/ComfyUI_LiteLLM, then pip install -r requirements.txt and restart.

Heads up: that requirements file is heavy - litellm, boto3, sentence-transformers (which drags in its own torch), plus a LightRAG fork. First install is slow. You don't need API keys for this node, but the moment you connect a completion node you will, and those go in environment variables (OPENAI_API_KEY, ANTHROPIC_API_KEY, …), never pasted into the graph - keys in widgets get embedded in the workflow JSON you share.

Where people trip

The usual mistake is chaining order - if you want system first, put that node at the start of the chain, because every node appends to the end. Also, nothing here validates your roles: you can send assistant as the last message, and the model will happily treat it as a nudge. That's sometimes exactly what you want (pre-filling the model's reply). It's a building block, not a nanny - and that's fine.