💬 Polza Chat

Most LLM nodes in ComfyUI want you to download weights and hand over some VRAM. Polza Chat goes the other way: it's a thin client for Polza.ai, an API aggregator that fronts 300+ chat models behind a single key. GPT-4o, Claude Sonnet, Gemini, DeepSeek, reasoning models like o1 and R1 - they're all a dropdown away, and none of them touch your GPU. The name is the point: this node doesn't run anything. It rents.

Why reach for it? The standard job is prompt enhancement - the LLM-in-the-graph pattern where a rough idea gets rewritten into a structured, model-appropriate prompt before it hits your encoder. Polza Chat is that, but hosted: you get genuine frontier-chat quality without owning the hardware or babysitting an Ollama server. It's also the natural front half of a chain - ask it to write an image description, feed the text output into Polza Text-to-Image, and you've built a tiny text-to-image pipeline out of two nodes.

How it works

The node builds a standard chat-completions payload and POSTs it to /v1/chat/completions on Polza's API. No streaming, no local model - just HTTP and an API key. Two details are worth knowing before you use it:

• The model dropdown starts at "Click Load models." The pack deliberately refuses to phone home during ComfyUI startup (set POLZA_FETCH_MODELS_ON_STARTUP=1 if you want it to). Instead, a "Load models" button on the node fetches the live model list from the /polza/models endpoint - which needs your API key set first. That trips people up on first install: no key, no list.
• It always re-runs. IS_CHANGED returns nan, so every time you hit Queue, the node calls the API again - even if nothing changed. On a per-call metered service that's a quiet way to burn money. Rerunning a workflow is rerunning a purchase.

The inputs that matter

• model - pick from the loaded list (default openai/gpt-4o). This is the one input you actually shop around in.
• prompt - your user message. Required.
• system_prompt - optional role=system instruction; the obvious place for "you are a prompt engineer…".
• reasoning_effort - none/minimal/low/medium/high/xhigh, for reasoning models. Set to none (the default) and it sends nothing; anything else sends a reasoning effort param.
• response_format - text or json_object. JSON mode is handy when the output feeds downstream automation, but treat the model's JSON as untrustworthy without validation.

The rest - temperature, max_tokens, top_p, the two penalties, seed - are standard chat knobs you can leave alone.

Outputs

Four: text (the reply, wires into any STRING slot), reasoning (chain-of-thought, when the model produces it), cost_rub (what the call cost, in rubles - Polza is a Russian-market service, so the currency isn't a typo), and total_tokens. The node's own widget shows a preview plus the token/cost line, so you don't need extra display nodes to see what you paid.

Installing it

Easiest via ComfyUI Manager: search ComfyUI-Polza → install → restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/PioneerMNDR/ComfyUI-Polza
cd ComfyUI-Polza
pip install -r requirements.txt

Then restart ComfyUI. Dependencies are just requests, Pillow, numpy, torch - nothing ComfyUI doesn't already ship, and there are no model files to download.

The API key comes from polza.ai, and you set it one of three ways: an api_key field on the node, export POLZA_API_KEY="pk-...", or a config.json in the node folder ({"api_key": "pk-..."}). If none are present the node fails gracefully with a "key not found" message in its own widget.

Common issues

"Click Load models first" - you clicked Queue before loading the model list. Set the key, click the button, then run. Errors like ❌ HTTP 4xx - the node surfaces API failures in its text widget, not the console, so look there first. And remember the actual limits: the API filters whatever the vendor filters - you can't abliterate a refusal you don't have the weights for - and your prompts leave your machine by design. It's MIT-licensed and open, so read the source before you trust it, which is decent advice for every node in the API-wrapper category. For a model you genuinely can't run locally, that's a fair trade; for a job your 8B local model handles fine, it's a tax.