☁️Mini Party

The full comfyui_LLM_party stack is genuinely a lot - API loaders, persona nodes, tool nodes, memory, RAG, loop controllers. Great once you need it, overkill when you just want to hand a string to GPT-4o-mini with an instruction and get a string back. Mini Party is the pack's answer to that: it's a whole LLM call - loader, prompt, model, credentials - collapsed into a single node.

What it actually does

Think of it as a text-transform function where the "function" is an LLM instruction. input_str is whatever text you're feeding in - a prompt fragment, a piece of data, a previous node's output. prompt is the instruction telling the model what to do with input_str; the default placeholder text, literally "input function here," is a hint about how the author intends you to use it - write it like you're describing a function's job, not like you're chatting. model_name picks the model, defaulting to gpt-4o-mini, which tells you the intended use case: fast, cheap, good-enough transforms, not heavyweight reasoning. The result comes back as output_str, ready to feed into whatever's next in your graph.

The optional fields are exactly what you'd expect for hitting an OpenAI-compatible endpoint: base_url and api_key for pointing at OpenAI directly, a relay, or a local server, and is_enable - this pack's standard kill switch - to bypass the call entirely without deleting the node. Leave base_url and api_key blank and, per the pack's config conventions, it'll fall back to whatever's set in config.ini rather than making you fill them in on every single node.

Why reach for this instead of the full LLM node chain

Everywhere else in this pack, one API call means several nodes wired together: an API-loader node, a persona/prompt node, a memory node if you want context, the actual invocation node, then Show Text to see the result. That's the right shape when you're building a real agent with tools and multi-turn memory. It's a lot of ceremony for "rewrite this caption to be punchier" or "extract the sentiment from this text." Mini Party skips straight to the call. If you find yourself duplicating the same three-node LLM chain across a workflow just to do small one-off text transforms, Mini Party is very likely the node you actually wanted.

Installing it

Same pack, same install path - there's no separate download for Mini Party:

• ComfyUI Manager: search "comfyui_LLM_party" in the node browser, install, restart ComfyUI.
• Manual: cd ComfyUI/custom_nodes && git clone https://github.com/heshengtao/comfyui_LLM_party, then pip install -r requirements.txt inside your ComfyUI environment, restart.

If all you're ever going to use is API-based nodes like this one, the pack's only_api branch (git clone -b only_api https://github.com/heshengtao/comfyui_LLM_party.git) is worth knowing about - it skips the local-model dependencies (transformers, llama-cpp-python) entirely and installs a lot faster, at the cost of losing the local-loader nodes. Make sure there isn't already a comfyui_LLM_party folder from a previous install attempt before switching branches this way; the README is explicit that two copies of the folder will conflict.

Common issues

Since credentials aren't required fields on the node itself, the most common gotcha is a silent auth failure: leave api_key blank on the node and forget to set it in config.ini, and you'll get an API error rather than an obvious "missing key" message. If model_name is something other than a standard OpenAI model - you're pointing at a relay or a self-hosted OpenAI-compatible endpoint - make sure base_url ends in /v1/, which is the format this whole pack's API layer expects; a base URL without the trailing /v1/ is a classic way to get a confusing 404 instead of a model response.

Keep the prompt instruction specific. Because input_str and prompt both flow into one call with no intermediate steps, a vague instruction gives the model nothing to anchor its output format to - you'll get a wall of prose when you wanted one line, or an explanation when you wanted just the transformed string.