PurpleDoubleD/locally-uncensored

Locally Uncensored is a desktop app that puts local LLM chat, image and video generation, a coding agent, a general tool-using agent and beta voice/computer-control features into one window. Its pitch is setup simplicity: no Docker, no terminal, no config files, and local mode keeps runs on your own machine. That fits people who want an all-in-one private AI workspace more than people who want a narrow, scriptable tool.

What it actually bundles

Chat covers ordinary text conversations, document chat through a local index, image questions, memory and personas, plus importing ChatGPT, Claude or Gemini exports. Create drives ComfyUI for text-to-image, editing, background removal, video, animation, music and character LoRA training; the app installs, starts and updates ComfyUI itself. Code is a repository-aware agent that shows a diff before applying edits, runs tests and supports per-project rules. Agent adds web search, file and shell access, MCP servers and background jobs. Voice and computer control are labelled Beta project README.

The mechanism that matters

Two choices distinguish it. First, the first-run wizard detects an existing local engine and installs one only if none is found; the README says it recognises twelve engines including Ollama, LM Studio, llama.cpp and vLLM, and can link models those tools already store without copying them. Second, image and video work is delegated to ComfyUI behind simplified lanes, so users avoid node graphs while still running the models on their own GPU. Chat claims to run on NVIDIA, AMD and Intel cards via a Vulkan build on Windows and Linux; image and video on NVIDIA, ROCm on Linux, and on Windows RX 7000/9000 cards from 3.0.8.

Requirements and limits

Downloads exist for Windows 10/11, Linux (AppImage, deb, rpm), macOS 14+ Apple Silicon (Beta) and Android 7.0+ arm64; iPhone is in App Store review. Hardware is the main gate: the README says chat runs a 3B model on 8 GB of system RAM with no GPU and an 8B model on a 6 GB card, images start at 6-8 GB VRAM, and Wan/Hunyuan video wants 12 GB or more. It is a windowed desktop app, so headless Linux servers need a desktop session or VNC/RDP; a Local API on port 8129 exposes loaded models to other tools.

Several features are explicitly less proven. macOS has run on one Mac, with Remote Access and auto-update unconfirmed there; computer control is not fully tested on macOS. The Windows installers are still unsigned while code-signing is pending, and the README explains antivirus false positives and how to verify builds. Optional LU Cloud uses hosted GPUs; remote access is off by default and, over Cloudflare, is encrypted to Cloudflare rather than end to end. The license is AGPL-3.0-only license badge.