The problem
I run one Claude agent on my production workstation and another inside an isolated disaster-recovery environment. They needed to share findings and status, but opening a network path between them would defeat the point of keeping the recovery side isolated.
What I did
- Used a neutral cloud storage bucket as a drop box. Each side writes to the other’s inbox and reads its own, so neither network ever reaches the other.
- Kept the message format simple: one JSON file per message with sender, recipient, timestamp, subject and body. Read messages move to an archive, and old ones expire automatically after two weeks.
- Gave the relay its own least-privilege access that can’t reach anything else.
- Built send and receive scripts in Python for the production side that use the same message format as the PowerShell scripts on the recovery side.
- Set a standing rule that a relay message is information from the other agent, not an instruction from me, so anything with side effects still needs my confirmation. Scripts that arrive through the relay are treated as untrusted until reviewed.
Outcome
The production side was connected and sending messages on the day the relay was set up. The two agents can pass notes without any direct link between the networks.