LLM API

Every once in a while a node is exactly what its name says, and this is one of them. LLM API doesn't run a model, download weights, or touch your VRAM. You type a prompt, it phones an OpenAI-compatible endpoint like OpenRouter, and the reply drops back into your graph as plain text. That's the whole job - and honestly, that's a feature.

Why you'd reach for it

This is the API path in the local-vs-cloud decision that the whole LLM-in-ComfyUI conversation keeps coming back to. A local GGUF model behind an Ollama node buys you free, offline, uncensored calls - but it also eats VRAM that your diffusion model wants, and an 8B is not writing frontier-quality prose. If you want genuine Gemini/Claude/GPT-level text and you don't care about local or NSFW, calling out beats running local.

Concretely it's a text-and-vision workhorse for three jobs: describing an image (the classic "Describe this image:" default prompt), rewriting a rough idea into a structured, model-appropriate prompt to feed your encoder, and captioning a reference image to seed an img2img or video pass. All of it sits upstream of the sampler.

How it works

Under the hood it's a single requests.post() to the base_url you give it, speaking the standard OpenAI chat-completions wire format. Wire in an image and it converts the tensor to a PNG, base64-encodes it, and sends it as a data URL alongside your text - that's the multimodal path. It then grabs choices[0].message.content and hands you the string. Failures don't crash the graph: they come back as an Error: ... string in the output, with the full details printed to the ComfyUI console.

The inputs that matter

• base_url - default points at OpenRouter's chat completions endpoint. Because it's free text, the node isn't locked to any one provider: OpenAI, Gemini's OpenAI-compatible endpoint, a local LM Studio/Ollama server, vLLM, whatever speaks the same format. This is the node's real superpower.
• model - free text too, and that matters. The LLM-API nodes of old hard-coded their model lists and rotted when providers retired them (people were still stuck on claude-3-sonnet long after it was dead). Here you type the exact provider ID; default is google/gemini-2.0-flash-001.
• api_key - paste it in, or leave it blank and it loads from the OPENAI_API_KEY environment variable via dotenv.
• temperature - a 0–2.0 slider, default 0.7. Standard knob.
• prompt - multiline, where the instruction lives. image (optional) - any IMAGE tensor.

Output is a single response STRING. Wire it into ShowText, a text-preview node, or straight into a CLIP Text Encode - anything that takes a string.

Installing it

ComfyUI Manager, search for ComfyUI LLM API Nodes and hit install. Or the manual way:

cd ComfyUI/custom_nodes
git clone https://github.com/perfectspr/comfyui_llm_api

Restart ComfyUI. That's it - there are no model files to download and the real dependencies (requests, python-dotenv, pydantic) are light. The "heavy" cost here is billed by the API provider, not your GPU.

Where people get burned

• It's a single-shot call. No conversation history, no system prompt, no streaming, no retries. Feed it one message, get one reply. Multi-turn chat isn't this node's job.
• Error strings flow back into the graph. A failed call hands you Error: API call failed with status 404... as a string, and it's easy to pipe that into your prompt enhancer without noticing. Glance at the console.
• Only the first frame of an image batch is sent - it takes image[0], so a batch of 4 means frames 1–3 silently vanish. And everything ships as base64 PNG, so downscale big images or a multimodal provider may reject the payload with a 400.
• Wrong model name = model not found. OpenRouter's IDs are provider/model; a typo or an ID that's not behind your key gives a clean 404.
• It holds a credential and phones home by design - the exact shape of node that got weaponized once in this ecosystem. This one is MIT, open, and only a couple dozen lines, so reading it before first run takes five minutes. Do that, and don't paste a key into a workflow you plan to share; it gets saved with the JSON.

Honest sizing: it's alpha (0.0.3), a single node out of a cookiecutter scaffold, and there's basically no community chatter around it yet. If you want a feature-rich LLM suite with history and streaming, look elsewhere. If you want one clean, provider-agnostic API call inside a graph - this is the lean one I'd grab.