Home

English · 简体中文

Wisp Science

An open-source, local-first desktop research workbench. You talk to a model; it reads and writes files in your project, runs Python and R in persistent sessions, reaches ~80 bioinformatics databases, and can push heavy work onto a GPU server over SSH — all from one conversation, with everything stored on your own machine.

Local-first means local-first. Projects, conversations, artifacts and settings live in a SQLite database on your disk. API keys and SSH passwords live in your OS keyring, never in that database. Nothing is uploaded anywhere except the model provider you configure and the databases you explicitly query.

Start here

If you want to…	Read
Install it and run a first conversation	Getting Started
Know what the agent can actually do	Built-in Tools
Fold a protein, review literature, build a figure	Skills
Query UniProt, PDB, ChEMBL, GEO, ClinicalTrials…	Bioinformatics Databases
Run on a GPU box, an SSH server, or WSL	Execution Contexts
Find your way around the window	Interface Guide

The four things Wisp adds to a chat window

1. It runs code that remembers. The python and r tools are persistent REPLs, not one-shot scripts. A dataframe loaded in one message is still in memory ten messages later. Each project gets its own environment, and each execution context (local, WSL, each SSH server) can point at a different interpreter.

2. It knows where your data lives. 42 bundled skills cover structure prediction, single-cell analysis, literature search, and figure composition. 24 bundled MCP servers put roughly 80 public biology and chemistry databases behind ordinary questions.

3. It can leave your laptop. Register an SSH host or a WSL distribution, probe it once, and the agent can submit long-running GPU jobs that survive a disconnect — with live progress and no model in the monitoring loop. See Execution Contexts.

4. It keeps a record. Generated figures, tables and files land in an artifact panel with the code that produced them. Starred items get immutable version history. Decisions, papers and data assets the agent records show up in a research graph. See Interface Guide.

Two ways to run it

• Desktop app (Tauri + WebView) — the full interface this wiki describes.
• Headless CLI (cargo run -p wisp-cli) — same agent loop, same skills, same Python/R REPLs, no window. Good for servers and scripted runs.

Project links

• Repository: https://github.com/xuzhougeng/wisp-science
• Homepage: https://wispscience.com/
• License and third-party terms: see THIRD_PARTY_LICENSES.md in the repo and in skills/

---

This wiki documents Wisp Science v0.23.0. Where a page describes a boundary or a known limitation, it says so explicitly rather than leaving it for you to discover.