Locally Uncensored

The only local AI app that does Chat + Agent Mode + Images + Video — all in one.

No cloud. No censorship. No data collection. Your AI, your rules.

Chat with AI personas, generate images, create videos — all running locally on your machine.

Getting Started · Portable Download · Features · Why This App? · Roadmap · Contributing

---

📸 Screenshots

Chat with Personas	Image / Video Generation
Model Manager	Create View with Parameters

---

v2.2.1 — Latest Release (Hotfix)

Fixed: Model unloading now works correctly, GPU offloading enabled by default (num_gpu: 99), silent error swallowing replaced with console logging. Resolves high CPU/RAM usage on systems with dedicated GPUs.

See the full Release Notes.

---

❓ Why Locally Uncensored?

Tired of switching between Ollama for chat, ComfyUI for images, and another tool for video? Frustrated with bloated UIs that need Docker and a PhD to set up?

Locally Uncensored is the all-in-one solution. One app. One setup. Everything local.

How it compares

Feature	Locally Uncensored	Open WebUI	LM Studio	SillyTavern
AI Chat	✅	✅	✅	✅
Multi-Provider (Ollama + Cloud)	✅	✅	✅	❌
Agent Mode (Tool Calling)	✅	❌	❌	❌
Model A/B Compare	✅	❌	❌	❌
Local Benchmark	✅	❌	❌	❌
Image Generation	✅	❌	❌	❌
Video Generation	✅	❌	❌	❌
Uncensored by Default	✅	❌	❌	⚠️
Memory System	✅	⚠️ (plugin)	❌	❌
Agent Workflows	✅	❌	❌	❌
LaTeX / Math Rendering	✅	✅	❌	❌
One-Click Setup	✅	❌ (Docker)	✅	❌ (Node.js)
25+ Built-in Personas	✅	❌	❌	⚠️ (manual)
RAG / Document Chat	✅	✅	❌	❌
Voice (STT + TTS)	✅	⚠️	❌	❌
Open Source	✅	✅	❌	✅
100% Offline	✅	✅	✅	✅

---

✨ Features

• Uncensored AI Chat — Run abliterated models locally with zero restrictions
• Multi-Provider Support — Ollama (local), OpenAI-compatible (OpenRouter, Groq, LMStudio, vLLM), Anthropic (Claude). Switch providers per conversation.
• Agent Mode (Beta) — Give your AI tools: web search, web fetch, file I/O, code execution, image generation. It chains tools autonomously to answer questions with real data. Best with Hermes 3.
• Model A/B Compare — Send the same prompt to two models side by side. Compare speed, quality, and token usage in real-time with parallel streaming.
• Local Benchmark — One-click benchmark any model on your hardware. Measures tokens/sec, time-to-first-token. Leaderboard across all tested models.
• Memory System — The AI remembers you across conversations. Context-aware injection scales with model size. Auto-extraction learns from every exchange. Export/import as .md or .json.
• Agent Workflows — Reusable multi-step agent chains. 3 built-in workflows (Research Topic, Summarize URL, Code Review). Visual builder for custom workflows with prompt, tool, condition, and loop steps.
• Image Generation — Text-to-image via ComfyUI with full parameter control
• Video Generation — Text-to-video with Wan 2.1/2.2, HunyuanVideo, LTX support
• Document Chat (RAG) — Upload PDFs, DOCX, or TXT files and chat with your documents
• Voice Chat — Talk to your AI with push-to-talk and hear responses with TTS
• LaTeX / Math Rendering — Full KaTeX support for inline and block math equations
• Chat Export — Export conversations as Markdown or JSON with one click
• Token Counter — Live token usage display with color-coded progress bar
• Keyboard Shortcuts — Ctrl+N (new chat), Ctrl+E (export), Ctrl+L (focus input), Ctrl+/ (shortcuts)
• Search Providers — Agent web search via SearXNG, Brave Search, Tavily, or DuckDuckGo. Configurable in settings.
• 25+ Personas — From Helpful Assistant to Roast Master, switchable via dropdown
• Model Manager — Browse uncensored + mainstream models. HOT/AGENT badges. Variant selector for multi-size downloads.
• Thinking Display — See the AI's reasoning in collapsible blocks
• Custom Dark Titlebar — No native Windows chrome. Sleek frameless window with drag region, matching the app's luxury dark aesthetic.
• Branded Installer — Dark-themed NSIS installer with custom header and sidebar images. Premium from first click.
• Linear/Arc UI — Compact, monochrome, collapsible settings. Premium feel.
• Privacy First — Zero tracking. All API calls proxied locally. No analytics.
• 100% Local — Everything runs on your machine (cloud providers optional)
• Standalone Desktop App — Tauri v2 Rust backend. Download the .exe, run it.

Tech Stack

• Desktop: Tauri v2 (Rust backend, standalone .exe — no Node.js required)
• Frontend: React 19, TypeScript, Tailwind CSS 4, Framer Motion
• State: Zustand with localStorage persistence
• AI Backend: Ollama (local text), OpenAI-compat (cloud/local), Anthropic (cloud), ComfyUI (images/video), faster-whisper (voice)
• Build: Vite 8 (dev), Tauri CLI (production)

---

🚀 Quick Start

Windows

git clone https://github.com/PurpleDoubleD/locally-uncensored.git
cd locally-uncensored
setup.bat

Linux / macOS

git clone https://github.com/PurpleDoubleD/locally-uncensored.git
cd locally-uncensored
chmod +x setup.sh
./setup.sh

The setup script automatically:

1. Checks for Node.js 18+, Git, and Ollama
2. Installs missing dependencies
3. Downloads a recommended uncensored AI model (~5.7 GB)
4. Starts the app in your browser

Manual Installation

Prerequisites: Node.js 18+, Ollama

git clone https://github.com/PurpleDoubleD/locally-uncensored.git
cd locally-uncensored
npm install
npm run dev

Open http://localhost:5173 — the app recommends models on first launch.

Image & Video Generation

No separate installation needed! When you open the Create tab:

1. The app checks for ComfyUI automatically
2. If not found, click "Install ComfyUI Automatically" — it clones, installs dependencies, and sets up CUDA in one click
3. Go to Model Manager → Discover → Image/Video and click Install All on any model bundle
4. Generate images and videos — everything is ready

The entire setup happens inside the app. No terminal commands, no manual config files.

One-Click Start (Windows)

start.bat

Launches Ollama + ComfyUI + the app in one go.

---

📦 Portable / No-Install

Don't want to install anything? Download the portable version — just extract and run. No admin rights, no installer, no registry entries.

Windows (Portable)

1. Go to Releases
2. Download the .exe installer
3. When the installer opens, select "Install for current user only (portable)"
4. Choose any folder (e.g., a USB drive) — done!

Alternatively, download the .msi for a traditional system-wide install.

Linux (Portable)

1. Download the .AppImage from Releases
2. chmod +x Locally-Uncensored_*.AppImage
3. ./Locally-Uncensored_*.AppImage

No installation needed — AppImage is portable by design.

macOS (Portable)

1. Download the .dmg from Releases
2. Drag to any folder (doesn't have to be Applications)
3. Right-click → Open (first time only, to bypass Gatekeeper)

Note: You still need Ollama installed for AI chat. The app will guide you through setup on first launch.

---

🧠 Model Auto-Detection

The app automatically detects all installed models across all backends — no manual configuration needed:

• Text models — Auto-detected from Ollama. On first launch, the app scans your hardware and recommends the best uncensored models for your system.
• Image models — Auto-detected from ComfyUI's models/checkpoints folder. Drop any checkpoint in there and it shows up instantly.
• Video models — Auto-detected from ComfyUI. The app identifies your video backend (Wan 2.1/2.2 or AnimateDiff) and lists available models automatically.

Just install models in the standard locations and the app picks them up.

🎭 Recommended Models

Text (Ollama) — scales to your hardware

Model	Size	VRAM	Best For
Hermes 3 8B	4.3 GB	6 GB	Agent Mode — uncensored + native tool calling
Hermes 3 70B	42 GB	48 GB	Best Agent — maximum power
Gemma 4 26B MoE	18 GB	8 GB	26B brain, runs like 4B. Vision + tools. Apache 2.0
Gemma 4 E4B	9.6 GB	6 GB	Vision + native tools. 128K context
Qwen3-Coder 30B	18 GB	16 GB	Best coding agent — 256K context, native tool calling
Qwen 3.5 Abliterated	5–18 GB	6–16 GB	Best overall intelligence
Qwen 3 8B Abliterated	5.2 GB	6 GB	Fast, great for coding
DeepSeek R1 (8B–70B)	5–42 GB	6–48 GB	Chain-of-thought reasoning
GLM 4.6 9B Abliterated	6 GB	8 GB	Strong coding
Llama 3.1 8B Abliterated	5.7 GB	6 GB	Fastest all-rounder
Llama 3.3 70B Abliterated	42 GB	48 GB	Maximum intelligence

Image (ComfyUI)

Model	VRAM	Notes
FLUX 2 Klein 4B	8 GB	Next-gen, fastest FLUX
FLUX.1 Dev / Schnell	8–10 GB	Best text-to-image
Juggernaut XL V9	6 GB	Best photorealistic SDXL
Pony Diffusion V6 XL	6 GB	Anime/stylized

Video (ComfyUI)

Model	VRAM	Output	Notes
Wan 2.1 T2V 1.3B	8–10 GB	480p	Best entry point, fast
Wan 2.1 T2V 14B FP8	12+ GB	720p	High quality
HunyuanVideo 1.5 FP8	12+ GB	480p	Excellent temporal consistency
LTX Video 2.3 22B FP8	16+ GB	480p+	Latest, fast inference

---

⚙️ Configuration

Environment Variables

Create a .env file (see .env.example):

# Path to your ComfyUI installation (optional)
COMFYUI_PATH=/path/to/your/ComfyUI

In-App Settings

• Temperature — Controls randomness (0 = deterministic, 1 = creative)
• Top P / Top K — Fine-tune token sampling
• Max Tokens — Limit response length (0 = unlimited)
• Theme — Dark or Light mode

---

🗺️ Roadmap

• RAG / Document Chat — Upload PDFs and chat with your documents
• Agent Mode (Beta) — Tool calling with web search, file I/O, code execution
• Memory System — Persistent agent memory with export/import
• Voice Chat — STT + TTS with local Whisper
• Multi-Provider — Ollama + OpenAI-compatible + Anthropic
• Model A/B Compare — Side-by-side model comparison with parallel streaming
• Local Benchmark — One-click benchmark with leaderboard
• LaTeX Rendering — KaTeX for math equations
• Token Counter — Live context window usage
• Chat Export — Markdown and JSON export
• Keyboard Shortcuts — Power-user shortcuts (Ctrl+N, Ctrl+E, Ctrl+/)
• Search Providers — Brave Search, Tavily, SearXNG, DuckDuckGo
• Regenerate / Edit — Edit user messages, regenerate AI responses
• Memory Debug — See which memories are injected into prompts
• MCP Support — Model Context Protocol for extensible tool calling
• Create Modes — img2img, upscale, inpainting, background removal
• Agent Workflows — Multi-step workflow chains with visual builder
• Plugin System — Extend the app with community plugins
• Mobile UI — Responsive layout for phone/tablet access

Have an idea? Open a discussion!

---

📁 Project Structure

src/
  api/            # Backend clients (Ollama, ComfyUI, providers, RAG, voice, agents)
    providers/    # Multi-provider system (Ollama, OpenAI-compat, Anthropic)
  components/     # React components
    chat/         # Chat UI, A/B Compare, Token Counter, Export
    create/       # Image/Video generation UI
    models/       # Model management, Benchmark, Discovery
    personas/     # Persona selection
    settings/     # App settings, Provider config, Search provider
    layout/       # App shell, Sidebar, Header, Shortcuts modal
    ui/           # Reusable UI components
  hooks/          # useChat, useABCompare, useBenchmark, useKeyboardShortcuts
  stores/         # Zustand stores (chat, model, settings, provider, compare, benchmark)
  types/          # TypeScript definitions
  lib/            # Utilities, benchmark prompts, chat export, model compatibility

---

🖥️ Platform Support

Platform	Status	Download
Windows (10/11)	✅ Fully tested	.exe / .msi
Linux (Ubuntu 22.04+, Debian, Fedora)	✅ Fully tested	.AppImage / .deb
macOS	🚧 Community testing	Build from source

Note: We actively test and support Windows and Linux. macOS builds are provided on a best-effort basis — we don't have Mac hardware for testing. macOS users can build from source (see below) and we welcome community feedback and PRs for Mac-specific issues.

Build from source (all platforms)

git clone https://github.com/PurpleDoubleD/locally-uncensored.git
cd locally-uncensored
npm install
npm run dev          # Development
npm run tauri build  # Production binary

Contributing

We welcome contributions! Check out the Contributing Guide for details on how to get started.

See our open issues or the Roadmap for areas where help is needed.

License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Built with privacy in mind. Your data stays on your machine. 🔒

If you find this useful, consider giving it a ⭐

Website · Report Bug · Request Feature · Join Discussion