Claude Science
Claude Science is an AI-powered virtual lab platform built with Claude Code and Opus 4.6.
Core trick
Agents with assigned roles hold turn-based rounds over WebSocket, and a code extractor saves executable snippets from the transcript as versioned artifacts.
Description
Claude Science is an AI-powered virtual lab platform built with Claude Code and Opus 4.6. It enables research teams — human + collaborative AI agents — to explore scientific questions, coordinate structured meetings, and generate reproducible research artifacts. Real research workflows are often fragmented across notes, tools, and code; Claude Science bridges this gap by letting users define agent roles (ML researcher, statistician, domain expert), conduct multi-stage collaborative sessions, and auto-extract executable code and documentation from agent discussions. Powered by Opus 4.6, it orchestrates long-context, sequential agent planning and execution with real file operations and exports (ZIP / notebook / GitHub). The demo shows agent selection, live agent dialogue, code synthesis, and artifact exports — giving users a reproducible starting point for real work. Website: https://claude-science.up.railway.app/
README excerpt
# Virtual Lab Web App (aka Claude Science) AI-powered research team collaboration platform. Create virtual research teams with AI agents, run collaborative meetings, generate executable code, and export results. Based on [virtual-lab](https://github.com/zou-group/virtual-lab) by Zou Group @ Stanford. ## Features - **AI Agent Teams** - Configure AI agents with specific roles (ML Researcher, Bioinformatician, Statistician, etc.) or use built-in templates - **Collaborative Meetings** - Agents discuss research problems in rounds, with real-time WebSocket streaming; start a meeting with only selected agents from the team page - **Code Generation** - Auto-extract code from meeting discussions into versioned artifacts - **Visual Editor** - Drag-and-drop agent graph with React Flow + inline prompt editing with Monaco Editor - **Multi-Format Export** - ZIP download, Jupyter/Colab notebook, GitHub-ready files - **Multi-User Support** - JWT auth, RBAC (owner/editor/viewer), team sharing - **Search & Analytics** - Full-text search, team stats, agent metrics, meeting comparison - **Onboarding** - AI-guided team composition with semantic stages and editable agent cards; see [docs/ONBOARDING_FLOW.md](docs/ONBOARDING_FLOW.md) ## Quick Start (Local Development) ### Prerequisites - Python 3.13+ - Node.js 18+ - (Optional) Docker & Docker Compose ### One Command Start ```bash cd local # First time only: set up backend + frontend cd ../backend && python3 -m venv venv && ./venv/bin/pip i