An always-on-top desktop AI assistant and agent harness supporting local/cloud multi-model orchestration, screenshot-driven interaction, and cross-device conversation continuity.
Xpdite is a desktop-first AI assistant and agent framework, currently at version 0.1.0 (Public Beta). Its core design philosophy is to embed AI deeply into desktop workflows rather than existing as a separate browser tab.
Interaction Design#
- Always-on-top floating window: Does not steal workspace focus, available at all times.
- Screenshot-first interaction: Alt + . hotkey captures screen and submits a query, enabling a "see-and-understand" zero-switch workflow.
- Multi-tab parallel contexts: Independent session states and per-tab queue management for quick context switching.
- Comprehensive execution timeline UI: Real-time rendering of model thinking steps, tool call progress, terminal output, and completion status.
Model Orchestration#
Manages local Ollama models and multiple cloud providers (Anthropic, OpenAI, Gemini, OpenRouter, LiteLLM) through a unified interface. Built-in GPU auto-detection and optimization: NVIDIA (nvidia-smi) → AMD (HIP_PATH) → CPU fallback, automatically setting Flash Attention, KV Cache q8_0, Keep Alive 30m, 4 parallelism, etc.
Agent Capabilities#
Supports multi-step execution with approval-aware operations. Integrates external tools via the MCP protocol (mcp[cli] SDK) while bundling inline tools including terminal, sub_agent, video_watcher, memory, and scheduler, registered via mcp_manager.register_inline_tools() and intercepted in the tool loop. Tool coverage:
- Web search and crawling: DuckDuckGo + crawl4ai + trafilatura + camoufox + nodriver
- YouTube video analysis: youtube-transcript-api for subtitles, yt-dlp for audio download, faster-whisper/whisperx for transcription
- Twitter integration: twikit
- Scheduled task automation: Built-in APScheduler
Cross-Device Continuity#
The Channel Bridge (TypeScript service) connects WhatsApp, Telegram, and Discord adapters, enabling conversation continuity between desktop and mobile messaging apps (not native mobile apps). Built-in notification events and storage system.
Content & Memory#
- Claude-style inline Artifacts: Models generate inline code/Markdown/HTML artifacts, persistable and searchable.
- Long-term memory system: Hybrid retrieval using sentence-transformers vector memory + BM25.
- @ file attachments: Instantly attach files via
@<file_name>in the input box. - Broad file format parsing: Code files, PDF (PyMuPDF), Office documents (python-docx/pptx, openpyxl), ODF/RTF (odfpy, striprtf), ZIP listings, image files.
Architecture Highlights#
Separated design with Electron + React frontend and Python FastAPI backend, communicating via IPC and WebSocket. Electron shows a Boot Screen on launch, passing startup progress via XPDITE_BOOT structured markers on stdout; the Python backend confirms readiness via HTTP health checks. Python ContextVars enable multi-tab state isolation, ensuring no state leakage between concurrent tabs. Data storage is based on SQLite, managed through DatabaseManager.
Supports Windows x64, macOS Apple Silicon, and Linux x64. End users can install via one-line scripts; developers need Bun, Python 3.13+, UV, and Git.
Unconfirmed Items#
- Hugging Face page: Mentioned in CLAUDE.md, but no explicit link found.
- Standalone documentation site: The
docs/directory in the repo contains architecture, API reference, configuration, and troubleshooting docs, but no independent doc site URL was found. - Channel Bridge configuration details: WhatsApp/Telegram/Discord pairing setup is in
docs/getting-started.md, not directly read. - Security details: Project has SECURITY.md but no independent security audit; local storage of cloud API keys and Channel Bridge message transport encryption are unconfirmed.