ComfyUI-Simple-LlamaCPP-Client

A lightweight custom node for ComfyUI that connects directly to a local llama.cpp OpenAI-compatible server.

It supports:

• 💬 Chat completions (/v1/chat/completions)
• 🖼 Vision models (image + prompt input)
• ⚡ Streaming mode (SSE token accumulation)
• 🧠 Clean Answer + Thinking separation
• 📦 Optional JSON-only output mode
• 🔑 Optional API key authentication
• 🎛 Minimal design (server-side parameters stay server-side)

---

✨ Features

✅ Works with any OpenAI-compatible llama.cpp server
✅ Supports system prompt + user prompt
✅ Optional image input for multimodal models
✅ Outputs: Answer, Thinking, JSON, Raw, Model Used
✅ Auto-detects model from /v1/models
✅ Clean dropdown UI (no ugly free-text params)

---

📦 Installation

cd ComfyUI/custom_nodes
git clone https://github.com/ai-joe-git/ComfyUI-Simple-LlamaCPP-Client.git

Restart ComfyUI.

---

🚀 Server Example

llama-server.exe ^
  -m Ministral-3-8B-Instruct.gguf ^
  --host 127.0.0.1 ^
  --port 8082 ^
  --mmproj mmproj.gguf ^
  -c 8192

---

🧩 Node Inputs

Input	Description
server_url	llama.cpp server URL (default: http://127.0.0.1:8082)
prompt	User message text
system_prompt	Optional system instruction
image	Optional IMAGE input (vision models)
api_key	Optional Bearer token

Model Selection

Input	Description
model_mode	Dropdown: auto / custom
model_override	Only used if model_mode = custom

Stop Control

Input	Description
stop_mode	Dropdown: none, preset:common_eot, preset:triple_hash, custom
stop_custom	Used only if stop_mode = custom

Text Cleanup

Input	Description
text_postprocess	Dropdown: fix_mojibake, none, ascii_quotes, fix_mojibake+ascii_quotes

(Default fixes Here’s → Here’s)

---

📤 Node Outputs

Output	Description
answer	Final cleaned answer
thinking	Reasoning if provided
json	Parsed JSON output (if enabled)
raw	Full raw server response
model_used	Model name used

---

📜 License

MIT License.