Collaborative computational notebooks with AI research agents
Built at Humans& Product Hackathon
Research is inherently collaborative, but the tools aren't. Computational notebooks today are single-player experiences — one person writing code, reading papers, and validating results alone. Scaling a research workflow means hiring more people and waiting for them to get up to speed.
Parallel turns your notebook into a multiplayer research environment where AI agents work alongside you as specialized team members. You set a hypothesis, and your team gets to work:
- Engineer — designs experiments, writes reproducible code, and executes it in sandboxed environments
- Researcher — searches arXiv and the web for relevant literature, surfacing papers and prior work
- Reviewer — pokes holes in your methodology, identifies biases, and flags statistical issues
Agents contribute findings directly into your notebook in real-time. You stay in control — review their work, run their code, and iterate together.
- Real-time collaboration — multiple users editing the same notebook simultaneously via Yjs
- AI agent roles — spawn Engineer, Researcher, or Reviewer agents scoped to your workspace
- Live code execution — run Python in isolated containers with GPU support (T4 through H200)
- Research tools — agents can search arXiv, browse the web, and execute code autonomously
- Rich output — inline plots, dataframes, images, and LaTeX math rendering
- Activity streams — watch agents think, plan, and contribute in real-time
apps/
web/ → Next.js frontend, Convex backend
agents/ → Cloudflare Durable Objects for persistent agent state
sync/ → Yjs WebSocket sync server (Hono + PartyServer)
sandboxes/ → Python execution containers (FastAPI + Modal for GPU)
packages/
agent-kit/ → Core agent types and tool implementations
agent-kit-react/ → React hooks for agent streaming and chat
ui/ → Shared component library
| Layer | Tech |
|---|---|
| Frontend | Next.js 16, React 19, Tailwind CSS, Monaco Editor, Tiptap |
| Backend | Convex (DB + auth + functions) |
| Agents | Cloudflare Durable Objects, AI SDK, Claude Sonnet 4.6 |
| Sync | Yjs, PartyServer, Upstash Redis |
| Execution | FastAPI, Modal (GPU sandboxes), Docker |
| Monorepo | Turborepo, Bun |
# Install dependencies
bun install
# Run all services in dev mode
turbo devIndividual services:
turbo dev --filter=web # Next.js app (port 3000)
turbo dev --filter=agents # Agent server
turbo dev --filter=sync # Sync server (port 8383)Built solo in 8 hours at the Humans& Product Hackathon.


