🐙 Kraken Ollama Prompt Chat

Prompt enhancement is the part of AI art that never stops being fiddly: you want a local LLM to turn "a sad robot in the rain" into something with lighting, lens, and mood - but you don't want to leave ComfyUI, and you definitely don't want to pay per request. 🐙 Kraken Ollama Prompt Chat connects you to your own local Ollama server from inside the graph, with a chat interface that holds a conversation, keeps a log, and hands the final text to your pipeline as a string output.

It's part of Kraken Tools (krakenunbound/comfyui-kraken-tools). It's the pack's "bring your own LLM" node: no API keys, no cloud, just Ollama running on a machine you control. This is squarely the 2026 prompt workflow - your prompt gets wrapped in a chat template and fed to an LLM anyway, so having an LLM in the loop to build it is the natural extension.

How it works

The node talks to Ollama through the official Python client. The inputs are simple:

• host - the Ollama server URL. Check this first. The default is http://192.168.1.93:11434, which is an IP address from the author's own LAN, not localhost. If you run Ollama on the same machine as ComfyUI, set this to http://localhost:11434. This is the single most common "why does it error" cause with this node.
• model - any model you've pulled, e.g. llama3.2. Small instruct models work fine for prompt writing and run on CPU if needed.
• action - the state machine: Connect (verify the server and list models), Send Message (send user_message and get a response), Clear Chat (reset the conversation), and Use Last Response as Prompt (push the latest AI reply to the prompt_output).
• user_message - what you send. The default is a decent template: "Create a prompt for: a futuristic cityscape at sunset."

The author ships a specialized system prompt aimed at AI art generation, so the LLM isn't just chatting - it's being steered toward art-prompt output. The conversation state lives in the node (it keeps current_host and the log between calls), which is why the workflow is action-driven: you Connect once, Send Message a few times to refine, then Use Last Response as Prompt to commit.

Outputs

• prompt_output (STRING) - the generated prompt. This is the one to wire into CLIP Text Encode or into the pack's Kraken Unbound Prompt / Kraken WAN Prompt Splitter.
• ai_response (STRING) - the full raw LLM response, in case you want to see everything it said before trimming it.
• conversation_log (STRING) - the whole chat history, for dumping to a text display or file.

Install

Install the pack as usual (ComfyUI Manager → search "Kraken Tools" → restart), then the extra step this node needs: Ollama's Python package. The pack lists it as an optional extra (ollama), so a bare pip install -r requirements.txt won't get it:

pip install ollama

Then make sure Ollama itself is running and has a model pulled:

ollama pull llama3.2

Then set the host input to http://localhost:11434 and hit Connect. If the node was registered before you installed ollama, restart ComfyUI.

Where people get burned

• The default host is someone else's LAN. 192.168.1.93 is the author's machine. Leave it and you'll sit on a timeout; change it to localhost:11434 and the node works. This is the #1 issue, by far.
• It's a chat interface, not a one-shot. You have to walk the action sequence - connect, send, then use-last-response. People who drag the node in and expect a single forward pass get confused at first. The log output makes the flow visible; use it.
• Async client + firewall. The node uses Ollama's async client. If it errors with a network/firewall message (the built-in troubleshooting text covers this), it usually means Ollama is bound to localhost-only or isn't running - not that the node is broken.
• Bigger isn't better. For prompt writing, a 3B-class model is faster and often writes cleaner art prompts than a big instruct model that wants to argue with you. Start with llama3.2 and only go bigger if the output quality isn't there.

If you've been tabbing out to a browser to ask an LLM to improve your prompt, this node removes that tab. Just change the host, or the first thing you'll see is a connection error named after someone's home router.