Building an AI Agent Mesh in My Homelab
How I built Hrolbot — a multi-agent AI system running on a Proxmox VM, where Claude Code agents coordinate through an MCP mesh to write and review real code.
A few months ago I had one of those ideas that sounds reasonable at 2 AM: what if I ran a fleet of AI agents on a Proxmox VM in my homelab? Not as a proof of concept, but as something I actually use daily. That idea became Hrolbot.
Hrolbot is a multi-agent system where each agent — a frontend developer, a backend developer, a reviewer — runs as an autonomous Claude Code instance. They coordinate through an MCP (Model Context Protocol) mesh, passing messages, sharing state, and handing off tasks to each other. An orchestrator agent receives instructions from Discord and delegates work to the right specialist.
The Proxmox VM handles the coordination layer: the MCP server, the message bus, the state store. The actual AI inference happens in the cloud via the Anthropic API, so the VM doesn't need to do any heavy lifting on that front. It just needs to keep the mesh running and the agents talking.
One of the more interesting challenges was getting the agents to work together without stepping on each other. Each agent operates in its own workspace, commits to its own branch, and communicates through structured messages. The orchestrator handles merge coordination and conflict resolution. It's not perfect, but it works surprisingly well for a system running on a 4-core VM with 8GB of RAM.
The whole thing is open source. If you're curious about running your own agent mesh — or you just want to see what happens when you give a VM more responsibility than it probably deserves — check out the repo on GitHub.