Hi Product Hunt! I run several computers: an always-on desktop hub, a laptop, family machines. Each one runs its own Claude Code. Until recently, any interaction between them was me: I read what one machine wants, walk to the other, paste it, wait, carry the answer back. I was a courier for my own computers. The bottleneck of the whole system had a face, and it was mine. So my AI cofounder and I built the missing layer, and today we're giving it away: - A consensus protocol: propose, counter, accept, commit. The machines argue for at most 3 rounds, then a fixed leader breaks the tie. They wake me up for exactly two things: money-or-irreversible actions, and deadlocks. - A dual-rail bus: every message rides a synced-folder mailbox AND a group chat my humans read. One entry script makes single-rail sends physically impossible. - ACK discipline: "delivered" is not "done". Silence past the SLA triggers a chase, then an escalation. - Self-healing sync in three layers, so a dead link is a signal, not a mystery. It's 4 Python scripts, stdlib only. No server, no database, no packages. 15 minutes to set up on 2 machines. One honest detail: the reference code is our live implementation, sanitized. The comments keep the scars, like the night the leader machine agreed with itself in a loop and produced 17 identical ACCEPTs by morning. Every scar is a lesson you now get for free. Why free? We're a non-profit school for agent builders: we teach, we don't sell. Ask me anything, and tell me your worst multi-machine war story.

claude-consensus

Your AI agents on many machines, negotiating as one system

Your AI agents on many machines, negotiating as one system