Grok 4.3

Yes, that's a Grok node in your ComfyUI graph. Runware's catalog isn't just diffusion models - it's an inference aggregator, and xAI's Grok 4.3 rides along as a text-generation node (xai:[email protected], textInference) that returns a plain STRING. If you've been building ComfyUI pipelines that need an LLM step - prompt enhancement, tagging, a caption that feeds back into a sampler - this is a way to get a frontier model into that graph without running it locally (which you basically can't) and without leaving ComfyUI for a browser tab.

The one thing that will trip you up immediately: the messages input is required and it's not a text box. It's a typed RUNWARE_MESSAGES socket, which means you build the conversation with the Runware/Params → Messages builder - one node per message, with role (user or assistant) and content (the text). Wire a Messages builder into the socket and you've got a conversation. This is the pack's whole design philosophy: stackable builders wire into model nodes, so a chat history is just a chain of message nodes. It's more nodes than a chat box, but it's also the same mental model as everything else in ComfyUI.

How it works

Standard pack plumbing: the node builds a request, sends it over REST through the runware-sdk, and joins any returned text into the text output as a STRING. The sampling knobs are real LLM knobs: settings.temperature (default 1) controls randomness, settings.topP (default 1) controls nucleus sampling, and settings.maxTokens is a gated toggle - flip it on to reveal the max-token value field. includeUsage will add token-usage stats to the response if you care about cost accounting.

The inputs that matter

• messages (required) - the conversation, built with Messages builders. Start with a user message.
• settings.temperature - determinism dial. Lower for reproducible outputs, higher for variety.
• settings.topP - diversity by probability mass; leave at 1 until you have a reason.
• toolChoice - gated; enable it to force tool calling, with toolChoice.type (strategy) and toolChoice.name (the specific tool). This is where you'd wire agent-style behavior.
• seed / numberResults - for reproducible or multiple answers.
• advanced_json - the escape hatch for tools (define function schemas as raw JSON merged into the request).

Output is text (STRING) - wire it into a Show Text node, or anywhere your workflow consumes a string. There's no native chat-streaming UI here; it's a request-and-return node.

Install and API key

Install once for the whole pack:

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

Restart ComfyUI (or install Runware from ComfyUI Manager). API key from runware.ai/api-keys, set in ComfyUI Settings → Runware API key, exported as RUNWARE_API_KEY, or via runware auth login.

Gotchas

The messages-builder requirement is where people get stuck, so repeat after me: no text box - build messages. Second, LLM calls on a metered API are billed by token, and a chat that runs in a loop (e.g., "improve this prompt, then pass it to the sampler, then critique the result") can quietly rack up spend; includeUsage exists precisely so you can see it. Third, settings.maxTokens is off by default, which is fine for short outputs but means a long-winded answer is governed by the model's default cap - if your responses get truncated mid-idea, that's the gate to open. Grok is a chatty model; budget accordingly.