The setup runs coding agents on disposable remote VMs while keeping personal credentials on a laptop. Herdr manages agents in persistent terminal sessions, shows whether they are busy or awaiting input, and lets agents start and review one another’s work. Ovis provides a phone and browser interface for transcripts and approvals, with notifications for questions that need attention. The security problem is that copying tokens to a VM exposes them to processes there, while ordinary SSH-agent forwarding gives the VM broad signing access without approval or a record of use.
The author’s tool, lend, forwards a policy-controlled proxy to the VM instead of the real SSH agent. When an agent requests a local command such as GitHub CLI, the laptop displays its executable, arguments, working directory, and stated reason; biometric approval runs it locally and returns its output, so credentials need not reach the VM. Policies can deny commands by default, restrict arguments and SSH keys, inject secrets from 1Password for approved commands, and grant short-lived, narrowly scoped leases. Lend also records activity without storing output or secrets, and can route other agent permission prompts to the laptop. Its limits matter: the SSH connection must remain active, request explanations from the VM are unverified, and an allowed command that reveals its environment or prints a secret can expose that secret to the agent.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.