A real LLM chat panel that lives inside ComfyUI — not a one-shot prompt box.
Running an LLM inside ComfyUI itself is slow and limited. So I ran mine in LM Studio / Ollama instead — but that meant constant alt-tabbing to copy prompts back and forth, and manually unloading whichever app I wasn't using so the other one had VRAM. And it was never just about prompts — sometimes I just wanted to ask a question, which meant opening a whole separate app for a five-second chat.
There wasn't a node that solved this properly, so I built one. Muse talks to the LLM server you're already running locally — no duplicate inference engine, no extra VRAM tax — so you get a real chat experience without ever leaving your workflow.
- Two backends — LM Studio and Ollama, switchable on the fly.
- Multi-session chats — create, switch, rename, delete. Each chat keeps its own history, system prompt, model, and settings, saved to disk and restored across restarts.
- Streaming replies with a Stop button, and a configurable max reply length so multi-item outputs ("write 5 prompts") don't get cut short.
- Hands-free VRAM coordination — Muse frees ComfyUI's VRAM before loading its own model, and confirms its own model is fully unloaded before your Run/Queue actually starts. No more remembering to unload anything yourself.
- Guide Materials — point Muse at
.txt/.md/.jsonfiles inComfyUI/input/and pick which ones shape every reply in a chat — your own style rules, official references, whatever you want the model to always follow. - Image attachments — drag-and-drop or pick images from
input/for vision-capable models, with no artificial per-chat cap. - Reasoning model support —
<think>output collapses into its own block, separate from the answer. - Per-message Copy / Regenerate / Delete, live token usage against the model's real context window, and a frosted-glass UI with an accent color picker.
It's a true drop-in: no pip install, no build step, no downloads. The only dependency is aiohttp, which already ships with ComfyUI.
- Copy this folder into your ComfyUI custom nodes directory:
ComfyUI/custom_nodes/comfyui-muse/ - Restart ComfyUI.
- Add the node: right-click the canvas → Add Node → utils → muse → Muse Chat.
- Pick a backend, confirm the base URL, hit ⟳ to list models.
- Type a message and press Enter. The model loads automatically and streams a reply.
- Drop reference files into
ComfyUI/input/and toggle them on under Guide Materials if you want standing context for that chat. - Copy whatever you like into your graph. That's the whole loop.
This node has no graph inputs or outputs by design. It's safely ignored by Queue Prompt — drop it into any workflow as a pure tool, not a dependency.
Two toggles in the top bar, both on by default:
- Free ComfyUI VRAM — unloads ComfyUI's models before the chat model loads.
- Unload on Run — unloads the chat model and confirms it's actually freed before your Run/Queue proceeds. Doesn't auto-reload — just send a new message when you want to chat again.
A manual Unload button is always available too.
Standing, per-chat reference material that shapes every reply — not a one-off upload.
- Drop
.txt/.md/.jsonfiles intoComfyUI/input/. - Check the ones you want active for the current chat.
- Muse reads them fresh from disk on every message, so edits apply automatically. Missing files are flagged, not silently dropped. Their token cost is included in the usage meter.
For vision-capable models — click + to pick from ComfyUI/input/, or drag-and-drop a file onto the panel (it gets saved into input/ first, then attached). Thumbnails persist in chat history; the actual image bytes are read live from disk each time, so chat files stay small.
- Iterating on prompts right next to the workflow that's going to use them.
- Cleaning up or expanding tag lists with a local model.
- Per-project chat threads with their own guide files — a house style that shapes every prompt automatically.
- Describing a reference image and asking for a prompt "in this style."
- Chatting freely, then just hitting Run — Muse gets out of the GPU's way on its own.
- A recent ComfyUI build (custom DOM widget support).
- A running LM Studio and/or Ollama locally, with at least one model.
- A vision-capable model for image attachments.
This is v1. If people find it useful, next up: broader local API support, MCP support, and whatever workflow gaps people actually report.
First public project — bug reports, feature requests, and PRs are all genuinely welcome.
Free, and staying free. If Muse saves you the back-and-forth it used to cost me, a coffee is appreciated: ☕ Support me on Ko-fi. A star or a share helps just as much.
MIT.