mflow syncs local worktrees so AI agents and humans stop overwriting each other. It runs a small daemon per worktree, watches file changes, and relays encrypted updates through a shared room. Without it, parallel agents drift silently until git finds the conflict too late.