Draggy

Draggy is a desktop AI assistant that talks, browses and works on your own projects, all on models running on your own computer.

It runs models directly on your hardware via its built-in native GGUF engine (powered by llama.cpp) with open weights from Hugging Face, so there is no account, no API key, and no request leaving the machine unless you ask for one. An Electron app, in React and TypeScript, for Windows, macOS and Linux.

The wiki, at draggy.org/wiki, is where the rest lives: installing, requirements, every feature, troubleshooting, and how the app is built, in the same twelve languages the app speaks. Start with Installation, or Building and architecture if you are here to work on the code.

• Chat with a local model. Draggy picks a model that fits your graphics card, downloads it, and lets you swap it any time. A Hugging Face browser lists the top 100 trending GGUFs, searches, and finds each quantization variant. Queue as many downloads as you like: two run at a time and the rest wait their turn. A download icon beside the search bar carries a red badge with the count, and opens a panel with each one's progress, time remaining (such as 2m 14s) and a cancel button that deletes the partial .download file.
• Stay quick on small models. An efficient system prompt keeps 1B to 7B models responsive without cutting any instructions.
• Keep chats and code apart. A switch at the top of the sidebar, and each side keeps its own history, model and settings.
• Work in your projects. Open a folder and it reads, edits and runs commands there, showing every change as a diff you can undo.
• Decide how much it may do. Four permission modes, from only proposing to working alone. Anything outside the mode asks first.
• Steer the plan. Tick, reorder and edit its checklist while it works, and let tasks carry on in the background.
• Search, browse and make files. Web search, a built-in browser that blocks adverts, and Word, PowerPoint, Excel and PDF files.
• Search your own documents. Point it at a folder, indexed on your machine and searched by meaning and keywords at once.
• Extend it. Forty-five MCP servers, the official registry and remote servers, plus 64 skills you start with / or leave to the model.
• Talk to it. A voice mode that knows when you have finished, answers out loud, and stops when you interrupt.
• Plug other tools in. An optional OpenAI-compatible API on 127.0.0.1, off by default and protected by a key.

Many QoL improvements, small features and changes.

Pick your system on draggy.org, or take the installer straight from the Releases page. Models are downloaded directly to your machine on first launch. See Installation for hardware requirements, the code-signing warnings each system shows once, and how updates arrive.

npm install

npm run electron:dev

Vite and Electron together with hot reload. See Building and architecture for how the code is laid out and where the two security boundaries are, CONTRIBUTING.md before sending a patch, and RELEASING.md for how versions are cut.

Recorded in app.log (sanitized operational events) and debug.log (raw diagnostic traces) in the application data folder. See Troubleshooting and Building and architecture in the wiki.

Tracked in the wiki's Troubleshooting page, along with what actually fixes each one.

GNU General Public License v3.0 or later. See LICENSE.

Draggy is free software: you may use, study, change and share it. If you pass it on, modified or not, you have to pass those same freedoms on with it and make your source available under the same licence. There is no warranty, to the extent the law allows.