Draggy: A Windows desktop app for local and cloud AI chat and coding

A local AI assistant and coding agent for Windows. Free, open source, no account.

Code mode on a sample project. Model waits are sped up 6 times.

Install from official sources only. Draggy comes from this repository, its releases page and draggy.org. Builds are not code signed, so third-party mirrors cannot be verified.

Draggy is a desktop assistant that talks, browses, and works in project folders, running open models on your own computer.

It runs weights locally using an embedded llama.cpp engine with GGUF weights from Hugging Face. There is no account, no telemetry, and no network traffic unless you request it. Built with Electron, React, and TypeScript for Windows.

Compared with LM Studio, Jan or Ollama, the difference is the coding agent: it reads, edits and runs commands in a project folder, shows every change as a diff you can undo, and works under one of four permission modes. Large Mixture-of-Experts models that do not fit in memory can still run from your drive.

Documentation lives on the wiki at draggy.org/wiki: installation details, hardware requirements, features, troubleshooting, and architecture in English, French, Spanish, and Chinese. Start with Installation, or Building and architecture to work on the code.

Guided first-run setup selects your interface language, theme, and an initial model without starting downloads until confirmed.

Run models locally through the built-in llama.cpp engine, with recommendations matched to your hardware.

Compact system prompts keep smaller models (1B to 7B parameters) responsive.

Sidebar modes keep general chat and project coding workflows in separate histories with independent models.

Project tools read, edit, and run terminal commands within a folder, with undoable diff reviews.

Four permission modes control tool autonomy: plan, ask, accept edits, and auto.

Interactive task lists let you check off, reorder, or edit steps during execution.

Web search and sandboxed browsing with ad blocking, plus document creation for docx, pptx, xlsx, and pdf formats.

Local document indexing and semantic retrieval over folders you select.

Extension support for 107 MCP servers, the official MCP registry, remote servers, and 66 built-in skills.

Large Mixture-of-Experts models that do not fit in memory run through a second engine, colibri, which keeps the always-used weights in VRAM and RAM and reads experts from your drive. Expect roughly 0.3 to 3 tokens per second on a good NVMe drive.

Voice mode with automatic turn detection, spoken responses, interruption handling, and an optional cloud voice.

Replies stream in word by word with a short fade-in.

Local image generation with /image or a tool call from the model. Stable Diffusion and Z-Image run on your graphics card, and a project gets the file saved in its folder.

Optional local OpenAI-compatible API on 127.0.0.1, disabled by default and key protected.

Optional third-party providers: connect local servers (Ollama, LM Studio, llama.cpp, vLLM, Lemonade) or hosted APIs (OpenAI, Mistral, DeepSeek, Groq, OpenRouter) with keys stored in the OS keystore.

Installing and requirements

Download the Windows installer from draggy.org or the Releases page. On first launch, Draggy suggests a model sized for your hardware and downloads it directly. See Installation for system requirements.

Image generation is optional and installed from Settings, under Models. The engine is a separate download of 17 to 900 MB depending on the build, and a model needs about 2 GB of video memory for Stable Diffusion 1.5 or 6 GB for Z-Image Turbo. With less, the weights stay in system memory.

Large models are managed from Settings, under Models, in the Large models panel. Draggy downloads colibri on demand, along with a small private Python runtime on Windows. Below 3 tokens per second it switches to slow mode: fewer tool rounds per turn, thinking off unless you turn it on, and a notification when a long answer is ready.

If Draggy is useful to you, a star on this repository helps other people find it.

npm install

npm run electron:dev

This starts Vite and Electron with hot reload. See Building and architecture for repository layout, CONTRIBUTING.md for contribution guidelines, and RELEASING.md for release workflows.

Operational events write to app.log in the application data folder, while detailed traces write to debug.log. Prompts and text payloads are hashed for privacy. See Troubleshooting for log locations across platforms.

Output quality depends on the selected model size and quantization.

Inference speed depends on available CPU and GPU memory bandwidth.

Large models depend on drive speed. A hard disk or USB drive is usually too slow to be useful.

License

Draggy is licensed under the Apache License 2.0. You can use, modify, share, and distribute it, including in commercial projects, provided you retain the NOTICE attribution file.

The license covers the code, not the Draggy name and logo. Provider logos shown in the app belong to their owners.