xuzhougeng/wisp-science

Search the literature, run Python and R, query ~80 scientific databases, and keep the trail — figures, runs, decisions, and drafts — in one project. Your data, conversations, and credentials stay on your machines.

Bring your own model. Keep your science.

What you can do

An agent that does the work

Bring OpenAI-compatible or Anthropic models, or drive Codex / Claude Code over ACP. The agent reads and writes project files, runs shell, and loads reusable Skills (SKILL.md) without flooding the prompt. Approval gates stay on unless you opt into Full Permission.

Compute from laptop to remote servers

Persistent Python and R kernels keep variables across cells and turns; each conversation gets its own isolated kernel, so parallel sessions never share state. Register local, WSL, and SSH hosts once; probe hardware; submit long Runs with live logs. Keys live in the OS keyring, never in SQLite.

Built for science

PubMed, GEO, and ~80 other databases through bundled MCP servers. Offline previews for notebooks, PDFs, Office files, and images. Isolated explorations to try a direction without touching the mainline. A Publication Workspace that freezes manuscript revisions into verifiable Evidence Capsules.

A workbench that remembers

Restart and the full history is back. Undo a turn’s file edits. Attach artifacts, files, and runtimes with @; search saved sessions with #; apply a skill with /. Encrypted manual sync and project transfer — nothing syncs in the background.

Get started

1. Open the download page, choose your operating system and processor, and download via Cloudflare. GitHub Releases is also available.
2. Open a bundled demo — no API key needed — to see a full RNA-seq trajectory.
3. Add a model in Settings → Models and start a project.

Setup walkthrough: Quick Start · basic configuration · model profiles · ACP agents

Build from source, CLI, and architecture: development.

Documentation

Community

Thanks to everyone who filed issues, sent PRs, and used Wisp on real projects. Special thanks to SpicyChicken6 for the molecular wordmark and three-wisp desktop icon designs.

Windows code signing by SignPath.io, certificate by the SignPath Foundation. Third-party notices live in development.

License

AGPL-3.0-only, except where a directory notes otherwise. Earlier releases keep the license published with them.

Citation

@software{xu2026wisp,
  author    = {Xu, Zhou-Geng},
  title     = {Wisp Science: a local-first AI research workbench},
  version   = {v1.5.0},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22009273},
  url       = {https://doi.org/10.5281/zenodo.22009273}
}