Parachute
A YHack 2026 bet that AI coding agents should be teammates you can watch work, not tools in another tab — so three of us built a browser IDE where people and agents edit the same files together, live.
Next.js 16Yjs / CRDTsWebSocketsMonacoFly.io
300ms
Sync debounce
3
Agent flow tiers
7 REST + 2 WS
Endpoints
What I built
Built a browser-based IDE where a team and its AI coding agents share one workspace — Monaco editor, a real terminal on xterm.js + node-pty, and live cursors showing who is editing what. Built by three people at YHack 2026 and deployed on Fly.io.
Kept concurrent edits conflict-free with Yjs CRDTs over WebSockets, then solved the harder half: two-way sync between the shared document and the real filesystem, where chokidar watchers, a 300ms write debounce and MD5 checks stop the two sides endlessly echoing each other.
Handled the failure mode agents add — two of them editing one file — with an orchestration layer that registers every agent, hands out file-level locks, and classifies each flow as primary, detour or integrator, so a colliding agent gets rerouted instead of causing a merge conflict.
Proved the loop with a hosted LLM agent (K2-Think-v2): full-workspace context, streaming over Server-Sent Events, typing generated code into the shared document character by character while everyone watches.