Wisp Science is a desktop and command-line workbench for scientific research that keeps papers, local data, code, and results inside one traceable project. Project files, conversations, and credentials stay on the researcher's machines; network calls happen only for the models, databases, agents, or remote compute the user configures.
The agent reads and writes project files, runs shell commands, and loads reusable Skills defined as SKILL.md files so instructions are pulled in on demand rather than carried in every prompt. Approval gates stay on unless full permission is granted, and a plan can be reviewed before it executes. Models are supplied by the user: OpenAI-compatible or Anthropic endpoints, or Codex and Claude Code driven over the Agent Client Protocol.
Compute is persistent and project-scoped. Python and R kernels keep variables across cells and turns, with one isolated kernel per conversation so parallel sessions do not share state. Local, WSL, and SSH hosts are registered once and can be probed for hardware, and long runs stream live logs. Keys are held in the operating system keyring rather than in the application database.
Around 80 scientific databases, including PubMed and GEO, are reachable through bundled MCP servers, and notebooks, PDFs, Office files, and images preview offline. Isolated explorations let a direction be tried without touching the mainline, and a publication workspace freezes manuscript revisions into verifiable evidence capsules. Tables, figures, code blocks, LaTeX formulas, and file paths become artifacts linked back to the conversation, code, logs, input files, and runtime that produced them.
Features
- Bring your own model: OpenAI-compatible or Anthropic endpoints, or Codex and Claude Code over the Agent Client Protocol
- Persistent kernels: Python and R state kept across turns, isolated per conversation
- Remote compute: local, WSL, and SSH hosts registered once, with hardware probes and long runs with live logs
- Scientific databases: roughly 80 sources such as PubMed and GEO through bundled MCP servers
- Research skills: reusable SKILL.md workflows applied with /, loaded only when needed
- Artifact provenance: figures, tables, and reports traced to their inputs, code, and environment
- Explorations: isolated branches for trying a direction without changing the mainline
- Publication workspace: manuscript revisions frozen into verifiable evidence capsules
- Session memory: full history restored after restart, with per-turn undo of file edits
- Local-first sync: encrypted manual sync and project transfer, with nothing syncing in the background
- Platforms: signed Windows MSI and NSIS installers, signed and notarised macOS .dmg for Apple Silicon and Intel, and .deb or AppImage for Linux on x86_64 and aarch64
