lend lets coding agents on a VM run commands on the owner's laptop after a Touch ID approval
The author of lend, which is published on GitHub as antoniomika/lend, runs his coding agents - Claude Code in one repository, Codex reviewing a branch in another - on virtual machines he can throw away, while his laptop keeps the SSH keys, the 1Password vault and the gh and cloud command-line tools logged in as him. In a post on his blog he describes lend, the tool he wrote so that those agents can use the laptop's credentials without receiving them.
He rejects the two usual answers. Copying a token onto the VM, for example into ~/.config/gh/hosts.yml, lets every process there read it for as long as it is valid. Forwarding the SSH agent with ssh -A keeps the private keys on the laptop, but the VM can use every key without asking and without any record of what was signed, and it only covers SSH keys.
lend is a proxy in front of the existing SSH agent; its socket is forwarded instead of the real one. SSH and Git keep working, and an SSH agent extension adds a way to run commands. When an agent on the VM asks to run gh, the command runs on the laptop with the owner's login and only the output, stderr and exit code go back. Every request, a signature or a command, is checked against a policy on the laptop that denies by default and lists allowed executables, optional argument prefixes such as pr view, and SSH keys by fingerprint. Approval can be a Touch ID touch inside the dialog. Agents reach it through lend mcp, an MCP server with one tool, run_local_command, whose reason field is required, or through lend exec on the command line.
A rule can also read a secret with op read and pass it to the approved command as an environment variable, so the token never reaches the VM. Leases allow, say, five minutes or three uses of exactly one approved command and disappear when the session ends or the screen locks; lend history records what ran without arguments, output or secrets. lend approve routes ordinary permission prompts from Claude Code, Codex, Antigravity and pi to the same dialog.
The author lists the limits himself. An approved command runs as his user and is not sandboxed; approving a shell, an interpreter or anything that reads stdin amounts to approving anything. The caller details shown in the prompt come from the VM and can be faked, and anyone who can reach the forwarded socket, root included, can make requests while the session is open, though the policy and the approval still apply. lend is early, runs on macOS and Linux, has no PTY, and biometric approval is macOS only.

Why it matters
The post moves the question from "which secrets does the agent hold" to "which commands may run, and who approved each one". The author's own warning sets the boundary: approval decides what runs, it does not contain what the command then does.