Kintsugi USB: A Multifunctional Rescue Boot Drive Integrating Offline Large Models and Agent Tools

Kintsugi USB is a Ventoy multi-boot USB solution based on Ubuntu 24.04. Its core features include integrating offline large model inference capabilities and mainstream Agent CLI tools, supporting one-click setup, and being suitable for AI-assisted programming and system rescue in offline environments. The project is maintained by jmagly, open-sourced on GitHub (link: https://github.com/jmagly/kintsugi-usb), and updated on 2026-06-13T15:15:16Z. Its name is derived from the Japanese kintsugi craft, embodying the design concept of integrating tools into an organic whole.

The name Kintsugi USB comes from the traditional Japanese repair craft 'kintsugi'. Its core philosophy is to integrate multiple independent tools into an organic whole, allowing developers to 'repair' and build software in any environment. Built on Ubuntu 24.04 LTS, the project uses Ventoy as the multi-boot management solution, eliminating the need for repeated USB formatting. It can update rescue images or add new system environments at any time, making it a 'living' tool drive.

The project has built-in offline LLM inference capabilities, integrating two major frameworks: Ollama and llama.cpp. Ollama provides a simple model management interface and supports one-click download and operation of mainstream open-source models like Llama2 and Mistral. llama.cpp is optimized for consumer-grade hardware, supporting CPU/GPU acceleration, allowing ordinary laptops to run models with billions of parameters smoothly. The offline feature ensures data privacy (sensitive code is inferred locally), making it suitable for enterprise developers, engineers in isolated environments, and traveling programmers, avoiding network delays and lags.

Pre-installed with mainstream AI-assisted programming tools, forming a complete Agentic CLI ecosystem:

• Claude Code: Anthropic's terminal AI assistant that understands codebase structures and performs multi-step programming tasks (modifying code, running tests, Git commits);
• Codex: OpenAI's code generation model that supports natural language-to-code conversion, suitable for rapid prototyping and completion;
• OpenCode: An open-source alternative that supports local deployment, aligning with the offline philosophy;
• GitHub Copilot CLI: Generates Shell commands from natural language, reducing the cost of memorizing complex commands;
• OpenClaw/omnius: Emerging Agent frameworks focused on multi-step task planning and tool invocation;
• Aider: A pair-programming AI assistant that collaboratively edits code files, supporting multi-file context and incremental modifications. The tools are carefully configured to work collaboratively, and developers can choose or combine them as needed.

It retains the core capabilities of traditional rescue boot drives, supporting rescue ISOs like GParted, Clonezilla, and SystemRescueCD through the Ventoy multi-boot architecture. The project provides a one-click setup wizard to simplify Live USB creation (executed with a single command). It is recommended to use a USB drive with a capacity of 64GB or more (to accommodate large model files), lowering the technical threshold and allowing more developers to enjoy the convenience of a customized Live USB.

It is suitable for various scenarios: a portable 'Swiss Army knife' for DevOps engineers (troubleshooting, deploying test environments); AI researchers can continue experiments in offline environments (planes, remote areas, confidential scenarios); educators can help students quickly get started with AI-assisted programming (without tedious configuration). Macroscopically, it reflects the trend of AI development tools becoming 'infrastructure'—large models are becoming basic components of development environments, and this project takes this concept to the extreme (packing the AI development stack into a USB drive).

Kintsugi USB represents a noteworthy paradigm for building development environments, leveraging the clever integration of existing tools to deliver greater value. The offline-first philosophy aligns with the awakening of data sovereignty awareness, and the pre-installed Agent tools showcase the latest progress in AI-assisted programming. For developers, it is an excellent entry point to learn skills like Ollama deployment, llama.cpp optimization, and Ventoy customization. Moreover, it inspires thinking: how to restructure workflows after the popularization of AI, so that technology serves people.