Llama Wrangler

Universal LLM Model Manager — download, convert, and hot-swap models with a single click.

Llama Wrangler is a cross-platform desktop app built on Electron that lets you browse HuggingFace and Ollama, download models directly to your machine, auto-convert them to GGUF format, and hot-swap between them — all without leaving the app. The app runs a local llama.cpp inference server on port 7070 and manages the full lifecycle of that server when you switch models.

Features

• Hot-swap models — switch the active model without restarting anything
• Embedded browser — browse HuggingFace and Ollama directly in-app; URLs auto-populate the download field when you land on a model page
• Smart downloads — detects pre-quantized GGUF files first; falls back to downloading + converting the base model if none exist
• In-app quantization — re-quantize any local GGUF to Q2_K through Q8_0 using llama-quantize
• Model management — view size, quantization type, and source location; delete models from the sidebar
• GPU auto-detection — uses Metal on macOS, CUDA on Linux/Windows when available
• Neo-Noir UI — frameless transparent window with a dark glassmorphism dashboard

Quick Start

git clone https://github.com/sanchez314c/llama-wrangler.git
cd llama-wrangler
npm install
./run-source-linux.sh    # Linux
./run-source-macos.sh    # macOS

See docs/QUICK_START.md for a step-by-step walkthrough.

Prerequisites

Dependency	Version	Required For
Node.js	18+ (22 recommended)	Running the app
Python	3.8+	Model downloads and conversion
huggingface-hub	latest	HuggingFace downloads
requests, tqdm	latest	Download progress
llama.cpp	built locally	Serving models

Install Python deps:

pip install huggingface-hub requests tqdm

The app will prompt to install llama.cpp on first launch if it is not found at ~/.llama-wrangler/llama.cpp/ or ~/.METALlama.cpp/.

Installation

From a pre-built binary

Download the release for your platform from Releases:

• macOS: Llama Wrangler-1.2.0.dmg (Intel, Apple Silicon, or Universal)
• Windows: Llama Wrangler Setup 1.2.0.exe or .msi
• Linux: Llama Wrangler-1.2.0.AppImage or .deb / .rpm

Build from source

npm install
npm run build          # current platform
npm run dist:mac       # macOS (Intel + ARM)
npm run dist:win       # Windows
npm run dist:linux     # Linux
npm run dist:maximum   # all platforms, all architectures

Built artifacts land in dist/.

Usage

Downloading a model

1. Navigate to a model page in the HuggingFace or Ollama browser tab — the URL auto-populates.
2. Alternatively, paste a URL or Ollama model name into the download field manually.
3. Click Download & Convert.

Supported inputs:

https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3
https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF
llama3.2
mistral
codellama

Switching models

Click any model in the sidebar. A dialog offers Load, Quantize, or Delete. Choosing Load stops the current llama.cpp server, starts a new one with that model, and polls http://localhost:7070/v1/models until ready (up to 30 seconds).

Quantizing a model

Open the action dialog for any local model and choose Quantize. Select a target level (Q2_K through Q8_0). The app calls llama-quantize and saves the output alongside the source file.

Server configuration

The llama.cpp server runs on port 7070 with:

• Context size: 8192 tokens
• GPU layers: 999 (auto-detects available VRAM)
• Host: 0.0.0.0 (accessible from localhost and LAN)

The server exposes an OpenAI-compatible API. Point any OpenAI client at http://localhost:7070.

Local storage

~/.llama-wrangler/
├── models/          # Downloaded GGUF files
└── llama.cpp/       # Auto-installed llama.cpp

Project Structure

llama-wrangler/
├── src/
│   ├── main.js              # Electron main process, IPC handlers, server management
│   ├── renderer.js          # UI logic, model list, download/quantize dialogs
│   ├── preload.js           # Secure IPC bridge exposed to renderer
│   ├── webview-preload.js   # CSS injection and security for embedded webviews
│   └── index.html           # Dashboard UI (Neo-Noir glassmorphism)
├── scripts/
│   ├── download_hf.py       # HuggingFace model downloader + GGUF converter
│   ├── download_ollama.py   # Ollama registry downloader
│   ├── compile-build-dist.sh
│   └── bloat-check.sh
├── resources/
│   ├── icons/               # .icns, .ico, .png
│   └── screenshots/
├── docs/                    # Full documentation suite
├── .github/                 # Issue templates, PR template, CI workflows
├── archive/                 # Timestamped backups
├── package.json
└── run-source-linux.sh / run-source-macos.sh / run-source-windows.bat

Troubleshooting

llama.cpp not found — the app shows a dialog to either select an existing build or install automatically. You can also point it at an existing build via the directory picker.

Download fails with "pip install" error — pip install huggingface-hub requests tqdm

Server port 7070 in use — check with lsof -i :7070 (macOS/Linux) or netstat -ano | findstr :7070 (Windows) and kill the conflicting process.

App won't launch on Linux (Permission denied) — run sudo sysctl -w kernel.unprivileged_userns_clone=1 or launch with --no-sandbox.

See docs/TROUBLESHOOTING.md for the full guide.

Contributing

Pull requests are welcome. See CONTRIBUTING.md for code style, commit format, and the PR process.

License

MIT — see LICENSE.

Acknowledgments

• llama.cpp for the inference engine
• HuggingFace for hosting models
• Ollama for model curation
• Electron for the cross-platform framework