H3 LLM Model (API)

H3 LLM Model (API) is a connection node, not a generator. Its entire job is to bundle up three config values - a base URL, a model ID, and an optional API key - into a single LLMMODEL output that the pack's H3 Story Director - LLM Model (API) consumes. Think of it as the adapter that lets the Director talk to whatever OpenAI-compatible text endpoint you happen to have, whether that's a local vLLM server, a company datacenter model, or a hosted inference service.

If Ollama is the "local and easy" path and OpenRouter is the "just works" path, this node is the "you have an endpoint already" path - the one for people who keep their own model server running and don't want a second vendor in the loop.

How it works

There's no magic here: the node holds three fields and hands them to the Director when it makes its chat-completions call.

• base_url - defaults to http://127.0.0.1:8080/v1. You can give it either an OpenAI-compatible base URL ending in /v1, or the complete /chat/completions URL. If your server isn't on the default port, this is the field you'll change.
• model - the model ID your endpoint serves. This is required in practice; the widget starts empty.
• api_key - optional for trusted internal datacenter endpoints. Only fill it when the server actually requires authentication.
• timeout_seconds - 300 default, up to 1800 for slow planning on a loaded model server.

The output is a single llm_model (type LLMMODEL), which plugs into the llm_model socket on H3 Story Director - LLM Model (API). The type is also compatible with nodes that accept YALLM's LLMMODEL type, so you can cross-wire it into the wider YALLM ecosystem if you've got one of those setups. On the flip side, you can feed the Director from YALLM's own LLM Model (API) / LLM Provider (API) outputs instead of this node - the socket doesn't care who made the LLMMODEL.

Install

Ships in the Vision Prompt Assistant pack:

cd ComfyUI/custom_nodes
git clone https://github.com/elgalardi/ComfyUI-VisionPromptAssistant

Restart ComfyUI. No models, no dependencies beyond the pack.

The security note

The api_key widget is visually masked, but the README is explicit that the value can still be serialized inside saved workflow metadata. If you're deploying workflows to a shared system, inject or strip credentials at deploy time rather than trusting the file. And if your endpoint needs a key at all, it's worth knowing the key stays in the graph - which is fine for a personal setup and a bad idea for one you post publicly.

Where it fits

Use this node when you already run an OpenAI-compatible model server and want the Director's full planning pipeline pointed at it. Keep in mind the model you serve needs to be multimodal if you're connecting reference images - the Director sends up to four images plus the JSON Schema and expects structured output back, and a text-only endpoint can't do that job. If you want a local multimodal model with almost zero config instead, the H3 Ollama Model (Local) node in this same pack is the friendlier option.