Clone Team Workspace

Multiplayer AI for teams. Every teammate is always in the room, even when they're not.

Track: YC Request for Startups, Fall 2026: Multiplayer AI — teammates join the same live AI session, see the work, steer it, and take it over.

Built at c0mpiled-18 · YC Startup School Hackathon, Taipei (Nov 11, 2026) on top of Clone's Team workspace.

AI today is single-player: everyone on a team has their own private chat, pastes context around, and nobody can join or steer anyone else's session. Clone already gives each person a Clone, a model of how they think, that predicts their next prompt. Team Workspace makes it multiplayer:

One live session	Teammates join the same AI session: presence, live cursors, one thread, one team plan.
See the work	The Team AI's replies and the team plan update on every screen the moment they change.
Steer it	Anyone can send the next prompt into the same session and change its direction.
Take it over	One click makes you the driver. The Team AI writes you a two-sentence handoff brief.
Pick whose Clone writes the next prompt	The composer predicts the next prompt. In a team you choose whose Clone predicts it: Clone Min, Clone Jun or Clone Sara each write it the way that person would, even when that person is away. Tab accepts, Tab again sends.

Run it

node server.mjs            # Node 20+, zero dependencies
open "http://127.0.0.1:4317/?as=min&ws=team"   # Min (CEO)
open "http://127.0.0.1:4317/?as=jun&ws=team"   # Jun (Engineer) in a second window
# Sara stays away; her Clone can still write the next prompt.

Model: any OpenAI-compatible endpoint, e.g. GMI Cloud inference:

LLM_BASE_URL=https://api.gmi-serving.com/v1 LLM_API_KEY=... LLM_MODEL=deepseek-ai/DeepSeek-V3.2 node server.mjs

Without those variables it uses the local Codex CLI signed in with ChatGPT.

How it works

• server.mjs: zero-dependency Node server. Realtime presence, cursors and room state over Server-Sent Events. One model call per prompt returns the Team AI reply, a team-plan item, and the next prompt each teammate's Clone would write. runPredict() writes those prompts on demand; takeOver() changes the driver and asks the model for a handoff brief.
• seed.mjs: the demo team and each Clone's memory of how its human thinks.
• public/: the room UI (vanilla JS). stage.html puts two teammates side by side.
• video/: record-real.mjs records the running app with two teammates (real model calls); public/motion.html + capture-mg.mjs render the demo frame by frame.

Demo video

video/clone-team-rooms-demo.mp4 (86 s, 2560x1440). A motion graphic rendered from public/motion.html. The predicted prompt, plan item and handoff brief shown in it are real model output from a take of the live app. Four segments labelled "Real app · live" (0:31, 0:53, 1:01, 1:13) are screen recordings of the running app with two teammates side by side; model wait time is cut.