xuzhougeng/wisp-science

Wisp Science is an open source desktop application that acts as an AI research assistant for scientific computing. It runs on your own computer rather than in the cloud, connects to OpenAI-compatible or Anthropic language models, and gives that AI persistent Python and R environments to work in, whether locally, over SSH to a remote server, inside Windows Subsystem for Linux, or on a machine with a GPU. Beyond running code, the app bundles around eighty bioinformatics and computational biology database tools that the AI can call through a standard called Model Context Protocol, so a researcher can ask questions that pull real data from these databases as part of a conversation. It also supports reusable Agent Skills, described in files named SKILL.md, which package up repeatable scientific workflows the AI can load and follow. The interface shows a live view of variables in the running Python or R session, similar to how RStudio displays its environment panel, and this view can be pinned as a separate window inside the conversation. Under the hood, the project is built with Rust for the backend logic, Tauri version 2 for packaging it as a cross-platform desktop app, and Leptos for the web-based interface shown inside that app. It also ships as a headless command-line tool for users who do not want the graphical interface. The codebase is organized into separate Rust crates handling the language model connection, the core agent loop and memory, file and shell tools, a local SQLite database for storing projects and conversations, the skills system, the runtime managers for Python and R, and the MCP client that talks to the bundled bioinformatics tools. Running it from source requires Rust, the uv Python environment manager, the Trunk WASM bundler, and the Tauri command line tool, with R support optional and never installed automatically by the app. Windows and macOS have prebuilt installers, while Linux users build from source. The README describes the current state as an MVP, meaning the core agent loop, providers, tools, and interface all work, but some planned features are still listed on a roadmap rather than finished.