LLM Config

LLM Config is the pack's "point it at an API and stop asking" node. Every LLM-powered node in this pack (the prompt enhancer being the main one) needs an llm_config before it'll do anything, and this is where that config comes from. Pick a provider, name a model, handle the API key - it packages all of that into a single LLM_CONFIG value you wire anywhere. Nothing runs locally, which is worth stating plainly: this is a key-management and routing node for cloud LLM APIs, not a model runner.

How it works

It's pure configuration logic, no network calls at all. Pick a provider and it fills in the known endpoint - OpenAI's api.openai.com/v1, xAI's for Grok, Google's for Gemini. The model string is whatever you want to use with that provider (defaults like gpt-4o and grok-2 are suggested). The api_key handling is the clever part and supports three ways in:

• Leave it empty - falls back to the provider's standard environment variable (OPENAI_API_KEY, XAI_API_KEY, GOOGLE_API_KEY)
• $ENV.VAR_NAME - reference any environment variable
• Type the key directly - works, but the code comments warn it can leak into workflow metadata, which lives in your saved PNGs

For a Custom provider, custom_endpoint lets you point at any OpenAI-compatible server - that's your ticket to local LLM backends like Ollama or vLLM, or any hosted clone. max_tokens caps the response (0 lets the API decide; some models require an explicit value).

The inputs

• provider - OpenAI, Grok, Google, or Custom
• model - the model name string
• api_key - direct key, $ENV.VAR_NAME, or empty for the default env var
• custom_endpoint (optional) - only used for Custom
• max_tokens (optional) - response cap, default 0 (API decides)

Output: llm_config, wired into any node that takes LLM_CONFIG input.

Install

Pack install via ComfyUI Manager (search "Latent Astronaut Suite") or:

cd ComfyUI/custom_nodes
git clone https://github.com/latentastronaut/comfyui-latent-astronaut-suite

Restart. On first boot install.py installs the openai and google-genai clients the LLM nodes need.

Where people get burned

Forgetting the key entirely is the classic failure - a blank key with no env var set gives you an auth error at generation time, not at config time. The $ENV syntax is the safer habit; hardcoding a key into the node is convenient but the pack's own docs warn it leaks into saved workflow metadata. And note the config node doesn't validate anything - typos in the model name or endpoint fail later, when the enhancer actually calls out.