Ugleh OpenAI - Responses Text

You've got a good idea for an image and a clumsy first draft of a prompt. This node is the fix: it ships your rough text off to OpenAI's Responses API and hands the polished answer back as a plain string you can feed straight into a CLIP Text Encode. One node, one output, no GPU touched, no local model to download. It's the minimal end of the "let a language model do your prompting" spectrum, and that minimalism is exactly its selling point.

It's the text-only sibling of ComfyUI's own official API nodes that landed for GPT-Image-1 back in 2025. Same idea - a cloud round trip inside your graph - minus the images. The tradeoffs are baked in: every call costs real money per token, it needs an API key, and you're talking to OpenAI's servers, so whatever they won't say, you won't get. If you want a private, uncensored local LLM, this isn't that. If you just want better prompts without leaving the graph, it's perfect.

The canonical workflow: rough idea into input, something like "Expand this into a detailed, comma-separated image prompt" into instructions, then wire text_out into a CLIP Text Encode's prompt. Instant prompt enhancement at generation time. It'll happily rewrite, translate, or brainstorm captions too - but rewriting prompts is the use case that justifies paying for a round trip.

How it works. Under the hood it's a thin wrapper around the openai Python SDK. Run the node and it loads your key, builds a client, and calls responses.create() with the model, your instructions as the system prompt, your input, the reasoning effort, and the token cap - then returns response.output_text. The pack keeps a tiny service layer so the node itself stays dumb. Nothing clever, which is the point.

The inputs that matter:

• model - a dropdown of ten models, from gpt-4o-mini up to the gpt-5 family. For prompt rewriting, gpt-4o-mini or gpt-4.1-nano is the sensible, cheap default. Break out gpt-5 when you need actual reasoning, and watch the bill.
• instructions - the system prompt, multiline, default "You are a helpful assistant." This is where you tell it what to do with your text.
• input - your text, multiline.
• reasoning_effort - low / medium / high. It steers how much the reasoning models think before answering; low is plenty for rewriting, high is slower and pricier.
• max_output_tokens - defaults to 300. Raise it for long rewrites, but every token is money.

The output is text_out, a plain STRING. Wire it into any node with a string input; CLIP Text Encode is the obvious target. If you want to eyeball the result before it touches a sampler, run it through a text display node first.

Install. ComfyUI Manager, search "Ugleh" and hit install - or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/Ugleh/ComfyUI-ugleh-OpenAI
cd ComfyUI-ugleh-OpenAI
pip install -r requirements.txt

then restart ComfyUI. Dependencies are modest - openai>=2.0.0 and python-dotenv, both legitimate PyPI packages. Then the key: create a file called .env in your ComfyUI root folder containing OPENAI_API_KEY=sk-..., or export it in your shell. No model files, no VRAM, runs on any machine ComfyUI runs on.

Where people get burned:

• Key not found - the node throws "OPENAI_API_KEY is not set." The pack's load_dotenv() reads .env from the current working directory, which is only your ComfyUI root if you launch from there. An exported environment variable is the more robust option.
• Stale openai SDK - the Responses API needs openai>=2.0.0. If you hit an AttributeError on output_text, an old SDK from another pack is shadowing it; pip install -U openai fixes it.
• Billing and access - a 404 on a model you picked usually means your account tier can't reach it, not a broken node. And don't leave gpt-5 at high reasoning_effort on autopilot; that bill sneaks up on you.

One aside: Ugleh is the same handle behind the QR-code-monster workflows and tutorials from 2023, so this is a veteran tinkerer's clean take, not a random script. It does exactly one thing, plainly. If you need vision, chat memory, or multiple backends you'll outgrow it fast - if you want a tidy text round trip, it's hard to beat. Usual custom-node rule applies: installing any node runs its code, so use Manager, and keep your API key out of any workflow you share.