ZenMux API 连接 / ZenMux API Connector

You don't need one node per LLM provider. ZenMux API 连接 / ZenMux API Connector is a single ComfyUI node for the ZenMux aggregator (zenmux.ai), which routes ~20 vendors - Anthropic, OpenAI, Google, DeepSeek, Qwen and friends - through one OpenAI-compatible endpoint. 138 models in the dropdown, and every single option carries its price in the label: anthropic/claude-sonnet-5 [入$2/M 出$10/M]. It's the most practical way to A/B the whole model landscape from inside a workflow without a stack of API keys and a stack of custom nodes.

That price-in-the-label detail is the killer feature, honestly. Picking a model becomes a budget decision you can read at a glance, and the list is sorted by vendor so the search box filters fast - type anthropic/ or qwen/ and the noise disappears.

How the model list works

Unlike the pack's YueGuang node, ZenMux's catalog comes from a bundled snapshot, zenmux/models_snapshot.json, so it stays current without a network fetch at runtime. When prices move, you refresh it yourself:

cd ComfyUI/custom_nodes/RUI-Nodes
python zenmux/build_snapshot.py

Old workflows survive updates: the node parses the model id out of a saved price-stamped label even when the price text no longer matches exactly, and VALIDATE_INPUTS is deliberately lenient so a stale label doesn't brick a whole graph.

Inputs and outputs

• api_key - ZenMux key. It saves into the workflow JSON, so scrub it before sharing files.
• model - default openai/gpt-5.4-nano, price-tagged and vendor-prefixed. Pick a vision-capable model if you plan to send images - otherwise they're silently ignored.
• system_prompt / user_prompt - role vs. ask; keep format rules in the system prompt.
• seed - realistically a re-run trigger; most of these models don't honor reproducibility anyway.

Optional: temperature (0.7), top_p (leave at 1.0), max_tokens (1024 - this is also your cost ceiling), image_1…image_6, detail (low/auto/high), image_max_size (default 1024; smaller = cheaper and faster), base_url (default zenmux.ai/api/v1, no :// - the frontend eats the protocol and the backend re-adds it), proxy_url (127.0.0.1:7890 style), usd_to_cny (display only).

Outputs: text, model_id (the real id, e.g. openai/gpt-5.4-nano, for downstream logging), and usage_stats - a four-line per-run report of tokens in/out, output character count, model + unit prices, and the USD/CNY cost conversion.

The retry magic

New models are fickle about sampling params. Some deprecate temperature outright; gpt-5 reasoning models insist on max_completion_tokens instead of max_tokens. Rather than making you chase per-model settings, the node catches the 400, strips or renames the offending parameter, and retries automatically. It logs the fix to the ComfyUI console - worth a glance if a call ever behaves unexpectedly, and it costs nothing on healthy requests.

Versus the YueGuang node

Same interface, same outputs, swappable. The differences that matter: ZenMux has the bigger catalog (138 vs 25 models) with vendor-prefixed ids and an online-refreshable snapshot; YueGuang keeps its list built-in and offline. If you're already on one aggregator, the other is one node swap away - which is honestly how this pack likes to do things.