Set Github Models LLM Service Connector 🐑

Of all the connector nodes in ComfyUI-MieNodes, this is the one you can try in the next five minutes without signing up for anything new - if you already have a GitHub account, you already have access to GitHub Models. It's Microsoft's marketplace-style gateway that lets you hit a whole spread of vendors' models (OpenAI, DeepSeek, Meta's Llama) through one endpoint and one token, instead of collecting a separate API key from every lab. This node is the plug for that gateway: wire its output into the pack's Translator or a prompt-generator node and you've got a working LLM path with zero new accounts.

Like the rest of the Set*LLMServiceConnector family here, it doesn't do anything on its own. It just resolves credentials and a model choice into an LLMServiceConnector object; the actual API call happens downstream, in whatever consumes that connector.

Inputs that matter. api_token is a GitHub personal access token - the README specifically recommends a fine-grained PAT rather than a classic one, which is the sane default anyway since fine-grained tokens can be scoped tightly. model_select gives you 12 options across three vendors: OpenAI's gpt-4.1 family (full, mini, nano) plus gpt-5-chat, gpt-5-mini, and o4-mini; DeepSeek's deepseek-v3-0324 and deepseek-r1-0528; and Meta's Llama 4 Maverick, Llama 4 Scout, and Llama 3.3 70B. Default is openai/gpt-4.1. If you want a model GitHub has added since this dropdown was last updated, pick Custom and type the exact model ID (vendor-prefixed, matching GitHub's own catalog naming) into custom_model.

One thing worth knowing before you lean on this for real work: GitHub Models is positioned as an evaluation and prototyping surface, not a production inference tier. Access is tied to your GitHub account and comes with its own rate limiting, separate from whatever the underlying vendor (OpenAI, DeepSeek, Meta) would give you on their own paid API. Great for testing which model actually gives you the prompt-rewrite quality you want before you commit to paying a vendor directly - less great as the backbone of a workflow you're running at volume.

The config-file gotcha, same as every connector here. prefer_local_config defaults to true: the node reads mie_llm_keys.json (copied from mie_llm_keys.json.example, kept next to the README, deliberately excluded from git) and, if the github_models entry there is non-empty, uses that key over whatever you typed into api_token on the node. This is the single most common source of "why is my key being ignored" across the whole pack - fill in the JSON file and leave the node field blank, or set prefer_local_config to false to make the node's own input win. config_key lets you point at a different entry if you keep multiple tokens.

Install. Search "ComfyUI_MieNodes" in ComfyUI Manager, or cd ComfyUI/custom_nodes && git clone https://github.com/MieMieeeee/ComfyUI-MieNodes and restart ComfyUI. No model downloads, no local dependencies beyond network access - this node is pure credential plumbing.

Troubleshooting. If requests fail immediately, double-check your PAT actually has the Models permission scoped in - a token generated for a completely different purpose won't automatically work here even if it's valid GitHub-side. If a specific model in the dropdown errors out, GitHub's Models catalog rotates faster than any node's hardcoded list can track, so custom_model with the current ID from GitHub's own Models marketplace page is your fallback. And remember this connector produces nothing to inspect by itself - if you're staring at a blank result, check that the output actually reaches a consumer node before assuming the connector is broken.