PurpleDoubleD/locally-uncensored

What it is

Locally Uncensored is a free, open-source (AGPL-3.0) desktop application for Windows, Linux, and macOS that bundles every major local-AI capability into one easy app. You install it, pick a model, and you're done — it handles the setup of the underlying AI engines (Ollama, LM Studio, ComfyUI, etc.) that normally require manual configuration.

The name comes from its positioning: it caters to users who want models that answer “directly, without refusals” (uncensored models), while still supporting mainstream models in the same interface.

Key capabilities

• Chat — Talk to local LLMs, upload images or documents for Q&A, voice conversations (beta, with 15 voices), memory across chats, personas, import ChatGPT/Claude/Gemini exports, and long chats get auto-summarized. "Uncensored" and standard models appear side by side.
• Create — Image and video generation on your GPU via a fully managed ComfyUI backend (no node graphs). The no-graph “lanes” handle text-to-image, image editing with a mask, remove background, text/video/animation, music, and Character Studio (train a character from your own photos into a LoRA).
• Code — An AI coding agent that maps the repo, shows diffs before applying changes, runs your testscher, and reads failures — with per-project rules via .lurules. Also exposes all local models via one OpenAI-compatible API endpoint for use with other coding tools.
• Agent — A general-purpose agent with tools for web search, file access, shell, code execution, screenshots, image/video generation, and your own MCP servers. Every tool call passes a permission gate; long tasks run in the background.
• Computer control (beta) — A model views the screen and uses the mouse/keyboard in apps you approve (off by default, requires explicit permission per action).
• Other — Remote access from your phone via LAN or Cloudflare tunnel, model comparison/benchmarking, cloud fallback (LU Labs, optional paid tier for models too big for a desktop GPU).

How it works

1. Flash your card — The app auto-detects any local engine already running (Ollama, LM Studio, Jan, vLLM, llama.cpp, etc.) or installs its own "LU Engine" with one click. It similarly installs/manages ComfyUI for image/video.
2. Pick models — A model manager marks what fits your hardware (VRAM) and downloads in one click.
3. Just use it — Chat, Create, Code, Agent are tabs in one window; you can keep the same model across tabs.

Tech & status

• Built with Tauri (Rust + web frontend), so it's a compact desktop binary.
• Current version: v3.0.6. Voice and computer control are beta; macOS supports Apple Silicon (beta). macOS builds are signed/notarized.
• Windows installers are unsigned until production signing is set up (free code signing via SignPath in progress); the update channel is signed with minisign.
• Curated model lists with VRAM requirements — e.g., Qwen 3.8 27B Uncensored at 11–18 GB, FLUX 2 Klein at 8–10 GB, Wan 2.1 video at 8–10 GB.

Bottom line

If you want a single appliance-style app for local LLMs, image/video gen, coding agents, and computer-use agents — without touching a terminal — this is the most complete straightforward option. The "uncensored" focus means it also curates models that avoid refusals, which you can either embrace or ignore.