RM-LLM 0.7.1

Nothing in RM-LLM 0.6.0 runs on your graphics card. It's an HTTP client wearing a ComfyUI node, and that decides everything else about it. You pick a provider, set a key, connect your prompts, and the request goes out to OpenAI, Claude, Gemini, Grok, DeepSeek, Mistral, Groq, Together, Fireworks, or the long tail via OpenRouter, Featherless and LithosAI - twelve providers, one node.

One thing to get straight before anything else. If your plan is a local abliterated 8B rewriting rough ideas into structured prompts, offline and free per call, that's the other branch of this pattern. Here you trade local for frontier-chat quality and zero VRAM - and you pay with a per-call bill, a filter you can't patch out, and prompts that leave the machine.

How it actually works

One class, RMLLM050, registered under RM/API. On load it builds its own optional sockets from cached provider capability metadata - OpenRouter's published parameter schema plus reviewed contracts for the nine newer providers - so the Advanced section changes depending on the model you've selected. Switch from a Claude model to a DeepSeek one and the setting sockets move under you. That's "model-specific controls", with a hard edge: connect a setting the model doesn't advertise and you get an error naming it, not a silently ignored parameter.

Requests go out as streaming chat completions, except Anthropic and Gemini, which use their native Messages and GenerateContent APIs. Reference media is inlined as base64, capped at 48 MiB, and video only goes out for OpenRouter and Gemini when the model advertises video input.

The fields worth touching

You can ignore most of the panel. provider and model_name are the two that matter - click model_name and the node fetches and searches the provider's catalogue, or type an ID if you're on Fireworks and need an account-specific model. api_key_env wants the name of an environment variable, never the key. credential_source chooses between that and a masked session key, and timeout_seconds defaults to 300. endpoint is the fiddly one: Auto lets OpenRouter route among the hosts serving your model and fall back when one is down, while pinning a host disables fallbacks entirely.

Then the optional sockets: image and video, console_output to stream text and reasoning into the ComfyUI log, creativity for a 0–100 slider that fans out to temperature and top_p together, sampling sockets like temperature and top_p where the model supports them, and input_budget / thinking_budget / output_budget with green usage bars after a run. key_ticket is session-key plumbing - leave it empty; a ticket is single-use.

Outputs are three strings: response, reasoning when the model returns any, and response_json for the full payload including usage and tool calls. Wire response into anything expecting a STRING - a prompt box, a save-text node, a second RM-LLM node.

Install

ComfyUI Manager, search the pack title, or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/RevenueMonkey/RM-Comfyui-LLM.git

Restart ComfyUI and refresh the browser; the node lives under RM/API. There are no pip dependencies to install and no weights to download - the project declares an empty dependency list and leans on what ComfyUI already ships. It does want a reasonably current ComfyUI, since it imports comfy_api.latest and the execution-context helpers. To update:

cd ComfyUI/custom_nodes/RM-Comfyui-LLM
git pull --ff-only

Keys: environment variable is the durable option - on Linux that's an owner-only file under ~/.config/rm-llm, and anything your launcher exports wins. Masked session keys live in server memory for that session. Either way keys are excluded from saved workflows, and masking is not encryption.

Where people get burned

Every queue is a real, billed request. Change-detection is deliberately disabled, so the node re-runs each time you hit queue even if nothing upstream changed. Correct for a text API, and a cost trap in a big graph.

A setting errors out or vanishes. You changed model and the new one doesn't advertise that parameter - or an endpoint pin removed it. The budget bars are estimates too.

Anthropic behaves differently. Its paired creativity slider is unavailable, so set temperature or top_p individually. Claude also needs an output token limit; the node supplies 4096 when you don't.

Your image was ignored. When a model doesn't advertise image input, the node warns and sends the request as text anyway, keeping the socket stable. New model names are assumed text-only until a capability record says otherwise.

Retries cover Featherless capacity and LithosAI rate limits; everyone else's 429 comes straight back at you. And 0.6.0 keeps the RM_LLM_050 ID so 0.5.0 workflows still load, but RM_LLM_040, RM_LLM_V3 and RM_LLM_V2 won't auto-convert.

One last thing, because this category earns it: a custom node is arbitrary Python with your user's OS access the moment ComfyUI loads it, and a node whose job is to hold your API key and call the network is exactly the shape of ComfyUI_LLMVISION, which shipped credential-stealing malware and ended in a federal guilty plea. This pack is new - first release late September 2026, no community track record - so read the code that touches your key.