← Back to the CV

Independent AI engineering · multi-agent platform

Guest agents, walled off

I run a multi-agent coding platform on one home server. Other people can use it too: each guest drives their own coding agents from Discord, on their own model key, pushing to their own git org. None of it can reach my home systems, my own agents, or the other guests. This page shows how those walls are built.

What a guest gets

Architecture

Guest agent architecture Guests chat with a separate guest orchestrator over Discord. The orchestrator and one worker container per project sit inside a firewalled network. Workers may reach the internet, the guest git forge and the publishing service. Worker-to-worker traffic, and anything toward the owner's agents, home systems, LAN, VPN or host, is dropped. Guests Discord · private channels Internet Model APIs · Discord · search Guest network · egress firewall Guest orchestrator Own Discord bot · own storage · own secrets Harness allowlist: git forge, web search, publish Encrypted key vault · usage ledger and budgets one container per project Guest A project 1 Guest A project 2 Guest B project 1 Guest git forge one org and bot per guest Publishing scoped by the guest's token ✕ worker to worker · ✕ worker to orchestrator Out of reach My own agents · home automation · infrastructure control The LAN · the VPN · the host itself
Everything a guest touches runs on one firewalled network. Solid arrows are connections the firewall allows; red crosses are the ones it drops.

The walls

What is still shared. Guests run on the same host and Docker daemon as my own agents. The guest orchestrator starts workers through the Docker socket, so it runs only code from my own images, and guest input reaches it as chat, never as code. The workers themselves have no socket.

Built with

Built and run through my own coding-agent loop, alongside the rest of the platform on the CV.