Hey, Dongle. 🔌

A Zero-Setup, USB-Portable Agentic Coding Assistant for Offline Environments

License: MIT | 100% FOSS | No Internet Required | Linux · macOS · Windows

The Problem

Over 2 billion people face unreliable internet monthly. But the deeper problem for developers isn't just connectivity—it's dependency.

Every modern AI coding tool assumes a working internet connection; a pre-configured machine, Ollama, or a runtime pre-installed; at least 16GB RAM; and a cloud API key. Students in rural India, engineers on air-gapped corporate networks, and developers at hackathon venues with broken WiFi—none of them fit that assumption.

Existing local tools like Goose, Cline, and Continue. Dev solved the privacy problem. They do not solve the accessibility and portability problem. They still require a full local dev environment just to run.

Hey Dongle solves a different problem: what if your entire AI coding assistant lived on a USB stick and just worked—on any machine—the moment you plugged it in?

What It Is

Hey Dongle is a self-contained, agentic AI coding assistant that runs entirely from a USB drive—no installation, no internet, no pre-configured environment required.

Plug it into any machine running Linux, macOS, or Windows with Python available, and within seconds you have a full agentic coding assistant that can read your codebase, write files, run code, and iterate autonomously.

It is not a chatbot. It is not a passive code explainer. It is an agent that takes instructions and acts on them.

How It Differs From Existing Tools

Hey Dongle vs Goose / Cline / Aider:

• Requires installation—Hey Dongle: None. Others: Yes.

• Requires Ollama or a runtime—Hey Dongle: Bundled inside. Others: Must be pre-installed.

• Works on CPU-only machines—Hey Dongle: Yes. Others: Partial.

• Works on low-RAM (8GB) machines—Hey Dongle: Optimized for it. Others: Struggles.

• Zero internet at any point—Hey Dongle: Fully air-gapped. Others: Setup requires internet.

• Target user — Hey Dongle: Developers anywhere. Others: Developers with good hardware.

Core Agentic Capabilities

1. Codebase-Aware Context

Hey Dongle reads your project directory on startup and builds a lightweight index of your files. It doesn't just answer questions — it knows what you're working on.

2. Autonomous File Operations

The agent reads files, writes files, creates new ones, and applies targeted edits—with your confirmation before any destructive action. This is a real agentic loop, not suggestions you copy-paste manually.

3. Code Execution & Iteration

It runs your code, reads the output or error, and iterates—up to N steps—until the task is complete or it asks for your input.

"Fix this bug" → runs code → sees error → edits file → runs again → done.

4. Natural Language Task Interface

Describe what you want in plain English. Hey Dongle breaks it into steps, executes them, and reports back. No memorizing commands.

5. Persistent Conversation Memory

Session history is stored locally on the USB. Come back the next day, plug in, and it remembers exactly where you left off.

Tech Stack

Model: Qwen2.5-Coder 3B (Q4_K_M GGUF) Best coding performance at small size. Runs on 6GB RAM, CPU-only; no GPU needed.

Inference Runtime: llama-cpp-python (bundled) CPU-first, zero-install, cross-platform. Ships inside the USB—nothing to download.

Agent Loop: Custom Python — no LangChain A minimal ~200 line tool-calling loop. Lean, fast, and fully auditable. No 40-dependency framework bloat.

Interface: Textual Python TUI terminal UI. No browser, no server, works on any machine.

Storage: SQLite (Python standard library) Session memory, file index, conversation history. Zero extra dependencies.

Packaging: PyInstaller + bundled model in a single folder on USB. One command to launch.

Hardware Requirements

Minimum:

• 6 GB RAM free

• Any x86_64 CPU (2015 or newer)

• 4 GB USB stick

• No GPU required

• No internet required

Recommended:

• 8 GB RAM or more

• 4+ CPU cores

• 8 GB USB stick

Use Cases

• 🌾 Developers in rural or low-connectivity areas

• 🏢 Engineers on air-gapped corporate or government networks

• 🎓 Students at institutions with restricted or limited bandwidth

• 🏁 Competitive programmers at venues with unreliable WiFi

• 🔒 Anyone who needs a fully offline, privacy-first coding assistant

Roadmap

Phase 1 — FOSS Hack 2026 (Current Ongoing)

• Zero-setup USB deployment

• Qwen2.5-Coder 3B bundled, fully offline

• Agentic loop: file read / write / execute / iterate

• TUI interface via Textual

• Persistent session memory via SQLite

Phase 2 — Post Hackathon

• RAG over local codebase for larger projects

• User-swappable model support (bring your own GGUF)

• Multi-file refactor planning

Phase 3 — Future

• MCP (Model Context Protocol) tool integration

• Support for even lower-spec hardware via further quantization

• Community model registry for offline download packs

License

MIT License — fully open-source, no proprietary dependencies, no cloud, no tracking.

Built for FOSS Hack 2026 🇮🇳