xuzhougeng/wisp-science

Open-source, local-first desktop AI research workbench for scientific computing with Python/R, MCP bioinformatics tools, SSH/WSL/GPU runtimes, and OpenAI/Anthropic models.

976 ★101 forksHTMLUpdated 21h ago

What you need to know

Open-source, local-first desktop AI research workbench and scientific computing environment (Rust + Tauri v2 + Leptos): connects to OpenAI-compatible and Anthropic models, runs persistent Python and R environments on local/WSL/SSH/GPU, loads reusable SKILL.md Agent Skills, and reaches ~80 bioinformatics databases through bundled MCP servers — data, conversations, and credentials stay on your own machines.

Install

Download installers from GitHub Releases: Windows MSI/NSIS, macOS .dmg (Apple Silicon + Intel)
Linux: build from source — Rust 1.88+ with wasm32 target, uv, Trunk, Tauri CLI v2; cargo tauri build
Headless CLI: cargo run -p wisp-cli (requires WISP_API_KEY / provider / model env vars)
Bundled bio-tools MCP: uv pip install mcp requests (plus server-specific deps like httpx, xmltodict)

Usage

  • export WISP_API_KEY=... WISP_PROVIDER=openai WISP_MODEL=deepseek-v4-pro; cargo run -p wisp-cli # interactive agent
  • cargo run -p wisp-cli -- run "Summarize the files in this project" (or --output jsonl for machine-readable events)
  • Launch bundled MCP bio-tools server via WISP_MCP_PKG=mcp_pubmed; agent discovers tools via search_mcp_tools / use_mcp_tool
  • Remote MCP (e.g. Notion): Settings -> Connections -> add Remote URL https://mcp.notion.com/mcp with OAuth
  • Optional ACP v1 agents: npm install -g @agentclientprotocol/codex-acp, then Settings -> Models -> ACP Agents

Key features

  • Agent loop (read -> think -> tool-call -> verify) with project-rooted path sandbox, approval gates, and opt-in per-conversation Full Permission mode
  • Persistent Python and R environments per project; local, WSL, and SSH/GPU execution contexts with runtime probing
  • ~80 bioinformatics databases (PubMed, GEO, ...) via bundled MCP bio-tools servers, discovered on demand instead of bloating every request
  • SKILL.md skill loading with progressive disclosure; drives external coding agents (Codex, Claude Code) over ACP v1 and controlled sub-agent delegation
  • Conversations persist to SQLite; one-click undo of a turn's file edits; fully offline previews for Jupyter notebooks, PDF, DOCX/XLSX/PPTX, images
  • Publication Workspace freezing manuscript revisions and exporting verifiable Evidence Capsules; encrypted manual sync and project transfer

Best for

Bioinformatics and computational-biology researchers who want a local-first, agent-driven workbench for literature search, Python/R analysis, and manuscript writing with bundled scientific MCP tools.

Caveats

  • Status: MVP vertical slice — agent loop, providers, tools, REPLs, SQLite store, MCP client, and UI build and run, but roadmap items are deferred
  • Windows/macOS installers are unsigned (SmartScreen 'More info -> Run anyway'; macOS right-click Open)
  • Linux requires building from source; R requires optional jsonlite package (never auto-installed)
  • Browser control only in Chrome/Edge environments
  • Licensed AGPL v3.0
Platforms: Windows · macOS · LinuxClients: Claude Code · Codex · Feishu · WeChat
Documentation ↗

Reviewed 2026-08-07

Topics

agent-skillsai-agentai-assistantai-for-scienceai4sciencebioinformaticscomputational-biologydesktop-appllmlocal-firstmcpmodel-context-protocolpythonreproducible-researchresearch-assistantrstatsrustscientific-computingscientific-workflowtauri
View on GitHub ↗
Stars
976★
Forks
101
Language
HTML
License
AGPL-3.0
Created
2026-07-01
Last push
2026-08-06