OpenAI API (Chat)

An OpenAI chat call as a graph node: give LF_OpenAIAPI a prompt, get back text - plus a cleaned, fence-stripped variant and a parsed-JSON variant, all ready to feed other nodes. It's the LF Nodes take on "LLM as a tool in the graph," for when you want GPT-class writing quality on a prompt, a caption, or some structured output without running a local model.

Same security-minded design as the pack's Gemini node: the node doesn't hold an API key. It posts to a proxy route on your own ComfyUI server, and the server forwards to OpenAI using a key it reads from its environment. Your key stays server-side and out of the workflow JSON you might share - which is worth more than it sounds in the wake of the ecosystem's LLM-node malware incident. Read the node's source before you trust it, but this pattern is the good one.

How it works

Required: prompt. Optional controls: model (default gpt-4 - point it at any model your OpenAI-compatible account allows), system_message (default "You are a helpful assistant."), temperature (0–2, default 0.7), max_tokens (default 1000), timeout (default 60s), and image (an optional reference image for multimodal models - wire in a render and ask GPT about it).

Outputs:

• text - the response text.
• clean - same text with code fences removed; the output to use when you asked for plain SVG or a bare prompt and don't want markdown scaffolding around it.
• raw_json - the full API response as JSON.
• json - parsed JSON when the response is valid JSON, else empty.
• image - the input image passed through, for multimodal flows.

The two "JSON" outputs matter more than they look: ask for JSON in the prompt, read it structurally via the json output, and you've got a mini-pipeline for structured generation.

Installing it and the key

One pack:

cd ComfyUI/custom_nodes
git clone https://github.com/lucafoscili/lf-nodes

restart ComfyUI, or ComfyUI Manager → "LF Nodes". The key goes to the server, not the node - set OPENAI_API_KEY (or OPENAI_API_KEY_FILE) in the environment before starting ComfyUI:

export OPENAI_API_KEY="sk-..."

Troubleshooting

• "No proxy URL configured" - the local proxy isn't reachable; restart ComfyUI with the key env var set.
• 401 / auth errors - the key is wrong or missing server-side. Check raw_json for the real message.
• Model not found - your OpenAI-compatible endpoint doesn't serve gpt-4. Point model at an ID your provider actually has. (The pack's README notes its LLM nodes were tested against KoboldCpp's OpenAI-compatible endpoint, so this pattern generalizes to local LLMs too.)