Wisp Science (Workspace for Intelligent Scientific Practice) is a local-first desktop research workbench in which an AI agent searches the literature, runs Python and R, queries scientific databases, and keeps the record of the work (figures, runs, decisions, and drafts) inside one project. Data, conversations, and credentials stay on your machines. It is a workbench rather than a model: you connect OpenAI-compatible or Anthropic models, or drive Codex or Claude Code over the Agent Client Protocol (ACP).
The agent reads and writes project files, runs shell commands, and loads reusable skills on demand, with approval gates on unless you opt into full permission. Persistent Python and R kernels keep variables across turns, and each conversation gets its own isolated kernel. Local, WSL, and SSH hosts can be registered once, probed for hardware, and used to submit long runs with live logs.
For scientific work it bundles MCP connectors with 247 tools across 23 connectors, reaching about 80 databases such as PubMed, GEO, UniProt, ChEMBL, PubChem, OpenAlex, and arXiv, plus 27 built-in skills. It previews notebooks, PDFs, Office files, molecules, and images offline. Exploration branches let you try a direction without touching the main line of a project, and a publication workspace freezes manuscript revisions into verifiable evidence capsules.
Restarting restores the full history, a turn's file edits can be undone, and projects can be synced or transferred with encryption on demand rather than in the background. It is built with Rust and Tauri and ships for Windows, macOS, and Linux, with bundled demos such as an RNA-seq analysis that run without an API key.
Features
- Research agent: reads and writes project files, runs shell commands, and loads SKILL.md skills, with approval gates
- Stateful Python and R: persistent kernels per conversation that keep variables across turns
- Remote compute: local, WSL, and SSH hosts with hardware probing and long runs with live logs
- Scientific connectors: 247 MCP tools across 23 connectors reaching about 80 databases, including PubMed, GEO, UniProt, and ChEMBL
- Built-in skills: 27 reusable analysis methods, such as single-cell quality control, clustering, and annotation
- ACP agents: drive Codex or Claude Code through the Agent Client Protocol
- Exploration branches: try an analysis direction in isolation from the project's main line
- Evidence capsules: a publication workspace that freezes manuscript revisions into verifiable records
- Offline previews: notebooks, PDFs, Office files, images, molecules, Markdown, and LaTeX
- History and undo: full session history after restart and per-turn undo of file edits
- Encrypted sync and transfer: manual project sync and transfer, with keys kept in the OS keyring
