ollama-usb

A portable, self-contained AI chatbot that runs from a USB drive on macOS. Plug in, double-click, chat — no installation required on the target machine.

What it is

• Local LLM inference from a USB stick — no internet, no cloud, no data leaves the machine
• Works on any Mac that has Ollama installed
• Clean browser-based chat UI with multi-model support and streaming responses
• One double-click to launch, Ctrl+C to shut down

Requirements

• macOS (Apple Silicon or Intel)
• Ollama installed on the machine you use for setup
• USB 3.0 drive (minimum size depends on models — ~10GB for small models, 32GB+ for larger ones)

Setup

Run this once on your own Mac to prepare the USB drive.

1. Clone this repo directly onto your USB drive:

git clone https://github.com/GhostInTheBus/ollama-usb /Volumes/YOUR-USB-NAME

2. Run the setup script:

cd /Volumes/YOUR-USB-NAME
bash setup.sh

The setup script will:

• Copy the Ollama binary from your local installation
• Show you a list of your installed models
• Let you choose which ones to copy to the USB

3. Eject and go.

Usage

1. Plug the USB into any Mac
2. Open the USB in Finder
3. Double-click launch.command
4. A browser window opens at http://localhost:8765 with the chat interface
5. Select a model from the dropdown and start chatting
6. Press Ctrl+C in the Terminal window to shut down

First run on a new Mac: macOS may block the ollama binary with a Gatekeeper warning. Right-click the ollama file → Open → click Open in the dialog. Then re-run launch.command.

How it works

USB Drive
├── ollama              macOS binary (copied from your Ollama.app)
├── launch.command      startup script
├── setup.sh            one-time setup script
├── models/             model files (not tracked in git)
│   ├── blobs/          content-addressed model weights
│   └── manifests/      model metadata
└── ui/
    └── chat.html       single-file chat interface

Launch flow:

1. launch.command sets OLLAMA_MODELS to the USB models folder and OLLAMA_HOST to port 11435 (avoids conflict with any locally running Ollama instance)
2. Starts ollama serve in the background
3. Serves chat.html via Python's built-in HTTP server on port 8765 (required to avoid browser CORS restrictions on file:// URLs)
4. Opens the browser

Recommended models

Model	Size	Good for
mistral:instruct	~4GB	General Q&A, fast
llama3.1:8b	~5GB	General chat, instruction following
llama3.2:3b	~2GB	Tight on space, still capable
qwen2.5:7b	~5GB	Strong reasoning

Pull models with ollama pull <model> before running setup.

Limitations

• macOS only — the included binary is macOS-specific. Linux and Windows support would require separate binaries and launchers.
• No autorun — macOS (and all modern OSes) block USB autorun for security reasons. One double-click is the minimum.
• Ollama must be installed on the machine you run setup on (to get the binary and models). The target machine does not need Ollama installed.
• RAM — model inference happens in the host machine's RAM. A 7B model needs ~8GB free RAM; a 30B model needs ~20GB.

License

MIT