A free, private, AI-powered desktop assistant that sees your screen. Take screenshots of anything, ask questions in natural language, and get instant answers -- all running locally on your machine with Ollama.

• Screenshot + Vision AI -- Capture any region of your screen (Alt+.) and ask questions about it
• Multi-Model Support -- Switch between any Ollama model from the UI; default: qwen3-vl:8b-instruct
• Streaming Responses -- Real-time token-by-token response display with thinking/reasoning visibility
• MCP Tool Integration -- Extensible tool system (file operations, web search, calculators) via Model Context Protocol
• Web Search -- DuckDuckGo-powered search and web page reading through MCP tools
• Voice Input -- Voice-to-text transcription via faster-whisper
• Chat History -- SQLite-backed conversation persistence with search
• Token Tracking -- Context window usage monitoring per conversation
• Stop Streaming -- Interrupt AI responses mid-generation
• Always-on-Top -- Frameless, transparent floating window that stays above all apps (including fullscreen)
• Mini Mode -- Collapse to a 52x52 widget when not in use
• 100% Local -- All processing happens on your machine. No data leaves your computer.

• Ollama -- Download from ollama.com and pull a model:

  ollama pull qwen3-vl:8b-instruct

Alternative: Download from the Releases page

Windows Security Notice: You may see a SmartScreen warning because the app is not yet code-signed. Click "More info" then "Run anyway".

1. Launch Xpdite
2. Take a screenshot with Alt + . (period)
3. Type a question or just press Enter
4. Get streaming AI responses in real-time

Watch on Youtube:

Step	Screenshot
1. Launch & capture
2. Enter a prompt
3. Real-time response
4. Final result

Electron (Desktop Shell)
    |
    +-- React Frontend (Chat UI, Settings, History)
    |       |
    |       +-- WebSocket <--> Python Backend (FastAPI)
    |                              |
    |                              +-- Ollama (LLM inference)
    |                              +-- SQLite (persistence)
    |                              +-- MCP Servers (tools)
    |                              +-- faster-whisper (voice)
    +-- Screenshot Service (Alt+. hotkey, region selection)

Tech Stack:

• Frontend: React 19, TypeScript 5.8, Vite 6, React Router 7
• Backend: Python 3.13+, FastAPI, Ollama, SQLite3
• Desktop: Electron 37+
• Tools: MCP SDK, DuckDuckGo Search, crawl4ai, faster-whisper
• Build: PyInstaller, electron-builder, UV

Xpdite uses the Model Context Protocol to give the AI access to external tools:

Adding new tools is straightforward -- see the MCP Guide.

# Clone and install
git clone https://github.com/KashyapTan/xpdite.git
cd xpdite
npm install
uv sync --group dev

# Run in dev mode (React + Electron + Python server)
npm run dev

# Build for production
npm run build
npm run dist:win

src/
  electron/           # Electron main process
  ui/                 # React frontend (pages, components, hooks)
source/               # Python backend
  api/                # WebSocket + REST endpoints
  core/               # State management, connections
  services/           # Business logic
  llm/                # Ollama integration
  mcp_integration/    # MCP server management
mcp_servers/          # MCP tool implementations
  servers/            # demo, filesystem, websearch
docs/                 # Documentation

Since the initial release, Xpdite has been significantly refactored:

• Complete backend refactor -- Modular architecture with separated concerns (api/, core/, services/, llm/, mcp_integration/)
• Model selection -- Switch between any installed Ollama model from the Settings UI
• MCP tool system -- Full Model Context Protocol integration with demo, filesystem, and web search servers
• Web search -- DuckDuckGo search and web page reading via crawl4ai with stealth mode
• Voice input -- Voice-to-text transcription via faster-whisper
• Thinking/reasoning display -- Collapsible display of model's chain-of-thought reasoning
• Tool call visualization -- UI cards showing tool executions with arguments and results
• Stop streaming -- Interrupt AI responses mid-generation
• Token usage tracking -- Real-time context window monitoring with visual indicator
• Chat history with search -- Browse, search, and resume past conversations
• Screenshot improvements -- Alt+. hotkey, fullscreen + precision modes, multi-monitor DPI awareness
• REST API -- Model management endpoints alongside WebSocket protocol
• Settings page -- Model enable/disable with toggle UI
• Conversation resume -- Full state restoration including thumbnails and token counts
• React hooks refactor -- Extracted useChatState, useScreenshots, useTokenUsage for clean separation
• Component library -- Modular chat components (ThinkingSection, ToolCallsDisplay, CodeBlock, etc.)
• Production docs -- Comprehensive documentation suite

See Contributing Guide for details.

MIT