VibeHQ Orchestrate multiple CLI agents as a real company team

I built VibeHQ because I got frustrated with "multi-agent" frameworks that are really just chain-of-thought with extra steps. VibeHQ takes a different approach: each agent is a real CLI process (Claude Code, Codex CLI, or Gemini CLI) running in its own terminal with its own working directory. They communicate through 20 MCP tools via a central WebSocket hub — not prompt injection.

Key ideas: - Contract-driven development: agents publish API specs and require sign-offs before coding begins. This eliminates the #1 failure mode — agents building against different assumptions. - Idle-aware message queue: messages queue when an agent is busy (detected via Claude Code's JSONL transcript), and flush when idle. No more "new instruction mid-task" corruption. - Full CLI preservation: every native feature (Skills, MCP servers, extensions) works out of the box. VibeHQ adds collaboration on top, never replaces. - State persistence: tasks, artifacts, contracts all persist to disk. Hub restarts don't lose state.

For the demo, I had 7 agents (PM on Codex, 5 engineers + 1 QA on Claude Code) build a full-stack hospital management system from a single PM prompt. The PM delegated tasks, engineers published specs, signed contracts, wrote code, and QA tested — all autonomously.

Demo video: https://drive.google.com/file/d/1zzY3f8iCthb_s240rV67uiA9Vps... GitHub: https://github.com/0x0funky/vibehq-hub

Currently developed/tested on Windows. Mac/Linux architecturally supported but untested (manual spawning works). Would love feedback on the architecture. The contract system and idle detection were the hardest parts to get right.