What is the problem your project solves?

AI agents are the most powerful new tool a team has — and the only one people still use alone.

A Claude Code run lives in a box only one person can see. The best you can offer a teammate is a read-only transcript they can't touch. As agents take on tasks that run for hours, that solo model breaks: work at that scale was never meant to be done by one person watching one terminal.

Docs had this problem until Google Docs. Design had it until Figma. AI hasn't had its multiplayer moment yet — that's the gap work-together closes.

How you are solving it?

One session = one agent = one cloud sandbox = one URL. Share the link and anyone can join from any device, see the same live transcript, and steer — type instructions, interrupt mid-run, hand off — while a live preview pane shows the app the agent is building.

How it works:

• Each session boots an E2B sandbox running a small Node bridge: a WebSocket server + a Claude Code agent loop.
• Browsers connect directly to the sandbox over WSS — no relay server in the middle. Late joiners get a full transcript replay; everyone gets token-by-token streaming.
• Sandboxes pause instead of dying, so a session resumes later with the agent's memory and transcript intact.
• A Next.js dashboard + Supabase registry maps session URLs to sandboxes.

Everything — bridge, protocol, UI — was built during the hackathon.

How Did You Use Claude?

Claude is the product's intelligence layer: every session runs Claude Code via the Claude Agent SDK (@anthropic-ai/claude-agent-sdk) inside the sandbox.

• Streaming input mode lets multiple humans inject messages into one live agent run — the core multiplayer mechanic.
• interrupt() lets anyone in the room stop and redirect the agent mid-task.
• Raw SDK messages stream over WebSocket to every connected browser for a live, token-by-token shared transcript.

We also built work-together itself with Claude Code.

What is the deployed URL for this project?

https://work-together-six.vercel.app