Claude Code on my laptop could read my folder of PDFs. Codex on my desktop needed one answer from those PDFs and couldn’t reach them.

So I copied the question from one terminal, pasted it into the other, waited, and copied the answer back. Again and again.

I was the slowest part of my own setup.

My agents didn’t need an orchestrator. They needed a mailbox.

My first instinct was to look for a framework. Something with task lifecycles, discovery, agents negotiating with each other.

Turns out I didn’t need any of that. Every agent belonged to me. None of them was a server. What they lacked was an address, and a reason to check for mail.

That second part is the one people miss. A coding agent runs a turn, then sits idle waiting for you. It never wonders whether a neighbour on another machine asked it something. If nothing nudges it, the message just sits there.

There’s a quieter trap too. Sessions are disposable. You /clear one, open another tomorrow, run two windows in the same repo. If you address messages to a session, mail goes missing all the time. The thing that stays put is the repo, so that's what I made the address.

So I built Kampung, a small self-hosted MCP server that gives every repo on every machine a mailbox. The name is the Malay word for village. Neighbours who can reach each other and lend a hand.

Two polite agents will reply to each other forever

The first real run worked. Codex, on a throwaway cloud box, found the right peer by itself, asked for my benchmark notes, and had written a test that reproduced the bug three and a half minutes after my one prompt. I never touched the Mac.

It also wasn’t tidy. Codex didn’t wait for the first answer and sent a near-duplicate. Then it asked for more, the Mac agent answered, and the two of them kept going until I stepped in.

Here’s the thing: every “thanks, got it” is another wake. Two helpful agents have no natural reason to stop.

My first guard counted how deep a reply chain went and cut it off after a few rounds. That stopped the loops. It also stopped honest back-and-forth reviews after about two rounds, which was most of what I actually wanted them to do.

The fix was to treat wakes as a budget that only a human can refill, not something the model can reset on its own. Mail still arrives past the limit. It just stops waking anyone.

Getting agents to exchange messages was one problem. Making sure each agent could only speak for its own repo was another. I nearly shipped a bug that let one window answer as a different repo’s mailbox, and a fresh code review caught it the day the branch was finished. I walk through that bug, how idle agents get woken, and how to try Kampung on one machine in a few minutes here:
→ How I Let AI Agents on Different Machines Work Together