SimpleCohereNode

ComfyUI is full of nodes that shuffle strings around: prompt boxes, wildcard pickers, text concatenators. SimpleCohereNode is the node that writes the string. It's a single chat call to Cohere's hosted Command models, dropped straight onto your canvas - no local model, no quantization, no extra gigabytes. If you've been eyeing the GGUF-based local LLM nodes but don't want to give up 6GB of VRAM to a 7B model, this is the lightweight alternative: the AI lives in Cohere's cloud, and your GPU stays free for the part it's actually good at.

The whole pack is one node, and it's genuinely simple. You feed it a system prompt and a user text, it calls Cohere's API, and it hands back a plain string you can wire into anything that takes text - most usefully a CLIP Text Encode prompt, so the flow becomes "sketch idea → LLM expands it into a real prompt → image." That's a clean loop to automate with a reroute, and it also quietly rephrases, critiques, or translates mid-workflow. The author's README is in Japanese and slightly out of date (it claims Command R+), but the shipped code pins command-a-03-2025 - Cohere's Command A - and that's what you actually get. README rot; code wins.

Mechanically there's nothing to overthink. The node reads your API key from the COHERE_API_KEY environment variable (with CO_API_KEY as a fallback), builds a system + user message pair, and sends it through Cohere's official cohere Python SDK. Two details worth knowing. First, it's an output node, which means a workflow "completes" when it runs - fine for prompt-generation workflows, just don't expect it to behave like a mid-graph helper. Second, and this is the one that'll bite you: every response is cached to a cache.pkl file inside the node's folder, keyed on system + text. Run the exact same prompt twice and the second run returns the cached string instantly, without spending API credits. Great for iterating on a workflow, mildly maddening when you change nothing and "fix" nothing - if an answer feels stale, delete cache.pkl or tweak the text.

The two inputs are all there is: system (default "You are a friendly AI assistant.") and text (default "Hello, how are you?"). Set system once to say who the model is pretending to be - "You are a prompt engineer for cinematic photography" beats the default by a mile - and put your actual request in text. The single output, result, is the model's reply as a STRING.

Installing is painless. Via ComfyUI Manager, search "comfyui_cohere" and install; or run:

cd ComfyUI/custom_nodes
git clone https://github.com/sugarkwork/comfyui_cohere

Restart ComfyUI. The only real dependency is the cohere Python package - no model files, no heavy wheels, nothing to download.

Where people get burned: you'll hit COHERE_API_KEY is not set the moment the node instantiates, because there's no UI field for the key. The README shows the Windows fix - add set COHERE_API_KEY=your_key to your ComfyUI launch .bat:

@echo off
set COHERE_API_KEY=COHERE_API_KEY
.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
pause

On Linux/macOS, export COHERE_API_KEY=... before launching works the same way. Grab a free Cohere trial key from their dashboard - it's enough to play with. The other trap is that the model name is hardcoded in the source, so if Cohere eventually retires command-a-03-2025 the node will start erroring until the author bumps it. It's a thin, single-purpose pack, but for "give me a good prompt without renting a local LLM," it does exactly one thing and does it cleanly.