Running coding agents on remote VMs keeps secrets off those machines, but agents still need to act as you: checking CI with gh, calling APIs with tokens, SSHing into hosts. The usual answers are copying credentials to the VM or forwarding your SSH agent — both give every process on the VM full, long-lived access.
The author's setup combines herdr, a terminal multiplexer that tracks which agent runs where and lets agents start and prompt each other, with Ovis, a browser and phone client that turns agent questions and approval requests into push-notification cards.
His new open-source tool, lend, addresses the credential gap:
- Agents run tools on his laptop through a proxy, with his explicit approval per request.
- Secrets and SSH keys stay on the laptop; the agent never sees them directly.
- Access is granted as revocable leases rather than permanent copies.
- He can see what the agent actually did and answer prompts from his phone.
The post also covers what lend does not protect against and prior art, and the tool is available on GitHub for anyone with a similar agent-on-VM workflow.
Summary by angular.love. The original source has the full details.