Ask Local LLM (GGUF)

Most "prompt enhancer" nodes only know how to be a prompt enhancer. Ask Local LLM is the opposite: it's the same local GGUF engine as the H3 Skill Promptor with the H3 contract ripped out, turned into a general-purpose chat node. Type a question, get an answer. Refine a prompt for any model, describe an attached image, brainstorm, summarize wired-in text - whatever you'd ask a frontier chat model, except this one runs on your card, offline, for free per call.

The LLM-in-the-graph pattern is older than this node and well-established, but what separates a good one from a toy is whether it keeps the VRAM cost sane and the output clean. This one does both by leaning on the shared server infrastructure from TrentNodes: it spawns (or attaches to) a managed llama-server and talks to it over OpenAI-compatible HTTP. No Ollama server to babysit, no separate process to configure.

The zero-reload trick

The defaults are identical to the H3 Skill Promptor's server spec field for field - port 8735, ctx 32768, mmproj auto-paired. The server manager only respawns when the spec changes, so if the promptor's Qwen3.8-27B is already resident, this node answers immediately, and vice versa. reasoning_effort is deliberately excluded from that comparison, so changing it never forces a reload. There's a real consequence here worth respecting: change ctx_size off the default and you force a cold respawn of a ~16 GB model. Leave it alone and the two nodes share one server like roommates.

What you set

• prompt and system_prompt - the system prompt is the only steering the model gets, so put your role and format rules there ("return only the prompt text", "answer in three bullets"). Default is a sensible direct-assistant.
• gguf_model / mmproj - from models/LLM; auto pairs the vision projector by filename prefix. Attach images and you're describing pictures, not just text.
• reasoning_effort - low / medium / xhigh, each with its own thinking allowance on top of max_tokens, so a long think can't starve the visible reply.
• seed - boring until it isn't: the node caches on identical inputs, so bumping the seed is how you get a different answer to the same prompt.

Three optional inputs earn their keep: input_text appends wired-in text under an INPUT TEXT header ("improve this prompt" flows wire straight in), history_json continues a previous conversation, and base_url points the whole thing at LM Studio, vLLM, or a llama-server you started by hand.

Multi-turn by wiring

The history_json output is the conversation as a readable JSON turn list. Wire it into another Ask Local LLM node's history_json input and you've got a follow-up with full context - which is how you'd build a chained critique-rewrite loop without any custom plumbing. Two guardrails keep that sane: history stays text-only (attached images become a [N image(s) were attached] note instead of re-sent base64, so chained turns don't balloon the context), and malformed history errors loudly rather than being silently dropped. The info output reports latency, token counts, and any warnings - worth a Show Text node if you're debugging.

Install

Same pack, same drill:

cd ComfyUI/custom_nodes
git clone https://github.com/TrentHunter82/TrentNodes.git
cd TrentNodes
pip install -r requirements.txt

or ComfyUI Manager → search "Trent Nodes" (if Manager flags it unsafe, that's the author's day-one repo-rename duplicate in the registry - the clone path works). Then drop a GGUF into ComfyUI/models/LLM - the tested pair is Qwen3.8-27B plus its mmproj from unsloth, but any current vision GGUF works. System prerequisite, once for the whole pack: a CUDA llama-server (llama.cpp ≥ b10450) on PATH, LLAMA_SERVER_BIN, or at ~/llama.cpp/build/bin/llama-server.

Where people get burned

• Attaching images without a vision projector raises immediately: put the mmproj next to the model, or remove the image input.
• Hitting the token limit mid-think returns an actionable error instead of an empty string - raise max_tokens or drop reasoning_effort.
• Leaked <think> blocks get stripped and reported in info; unlike the promptor, markdown fences are kept here, because in a chat answer they're usually intentional code.

The honest caveat: an 8–27B local model isn't writing better English than a frontier API, and you should not expect it to. What it buys you is uncensored, offline, free, and wired directly into your graph - and for the "translate this idea into a structured prompt" job, that's the right trade.