QuixiAI/QuixiDesk

QuixiDesk is a desktop AI workspace: a native tao + wry shell where you attach local folders, grant file-modification permissions, and hold multi-turn streaming sessions with a chosen LLM provider. It is aimed at people who want a general chat/task client with local file and browser context, rather than at developers building agents.

What you get in one window

The application bundles a Provider Manager and an MCP host. The engine routes across cloud, OAuth, and local providers: OpenRouter, custom OpenAI-compatible endpoints, OpenAI, xAI, Anthropic, plus Gemini, DeepSeek, Groq, Mistral, Together, Fireworks, Perplexity, Cohere, and Cerebras, and auto-detected local Ollama, LM Studio, or llama.cpp/LocalAI servers (README). Any OpenAI /v1/chat/completions endpoint can be plugged in, so a self-hosted model is an option without a separate client.

Alongside chat it ships pre-configured tools, including a workspace file indexer, git pool, shell host, AppleScript, PDF, memory, fetch, GitHub, Office 365, and scheduled tasks. A docked in-app browser provides tabbed navigation and [ref_N] accessibility element inspection, and macOS native computer use adds CoreGraphics screen capture, cursor/keyboard control, and window inspection. The README states 44 tools are active out of the box.

Platform reality

The core shell builds on macOS, Linux, and Windows using each system's webview, with a stable Rust toolchain and no Node.js or Electron for the production app. However, native computer-use integration (CoreGraphics/AppleScript) and the folder picker are macOS-specific; Linux and Windows builds do not have feature parity (README). macOS 14 or later is documented for the build/test setup. Linux needs GTK 3 and WebKitGTK 4.1 runtime libraries and an active graphical session, not a plain SSH shell; Windows needs the MSVC Rust toolchain, Visual Studio C++ Build Tools, and WebView2, with no MSI or signing workflow yet. Builds must be made on the target OS.

Caveats worth knowing

• Scheduled tasks can be saved and run manually, but automatic dispatch is not implemented.
• Interface: the replacement interface is currently English-only.
• Tests: the full E2E packaged-app runner is designed but not implemented; manual desktop checks remain separate from the automated regression and storage tests.
• Data safety: conversations save atomically and fall back to a recovery/ snapshot; if both save locations fail, quitting is blocked until a durable save succeeds, though a force kill can still lose in-memory data.
• Licensing: MIT (license).

The project asks for credit to the REA reverse-engineering tool used in its native-runtime analysis, which is unrelated to end-user functionality.

HotRepo's take

HotRepo recorded 107 total stars at #15 on its NEW-REPOSITORY Rust board on 2026-10-29, but has no daily-board history for this project. That single snapshot says nothing about sustained or fading interest; judge fit from the platform gaps and missing scheduled automation instead.