AI NewsDev toolsAnnouncement

Antonio Mika built Lend to keep agents away from your SSH keys

Antonio Mika published Lend. A coding agent on a remote server can use your SSH keys without the keys ever leaving your laptop.

AI News

Editorial2 min read

LinkedInX
Signal field artwork for a security tool that keeps credentials off remote coding agents

Why it mattersA team that wants coding agents on a server can stop copying keys across. Each sensitive call waits for a Touch ID press on the owner's laptop.

A developer runs Claude Code on a remote server. The agent now needs the developer's SSH key to push a branch. The simple answer is to copy the key onto the server. Antonio Mika wrote a tool called Lend that keeps the key on the laptop. Each use asks for a Touch ID press first.

Mika published Lend on 13 November. He posted it to Hacker News on 16 November. It reached 82 points in under five hours. The code is on GitHub at github.com/antoniomika/lend.

How Lend works

Lend sits in front of the SSH agent on the laptop. The agent on the server sends each credential request back through SSH to Lend. Mika writes that Lend checks the request against a policy file. It then shows the command to the user. The user approves. Only then does Lend run the command on the laptop. The output goes back to the remote agent. The key itself never leaves.

Approval happens through Touch ID on macOS. On Linux it uses a Zenity dialog. The policy can allow some commands without a prompt. So routine reads do not need a tap every minute. A separate MCP server exposes the same approval flow to coding agents that speak the protocol. Lend can also inject secrets from 1Password into one-off commands.

What Lend does not do

Mika lists the limits himself. Lend does not allocate a terminal. So an interactive program like psql or a text editor cannot run through it. Biometric approval is macOS only. Linux has no standard for it. Any process on the remote machine that holds the forwarded socket can send a request. A root user on the server can still spoof the caller name in the policy check. And once a command is approved, it runs with the user's full permissions on the laptop. Mika warns against approving bash, sh or any program that reads its own instructions.

He also says the software is early. He is its main user so far.

Who this helps

A team running agents on a shared server has one working pattern today. Give the server its own credentials. Scope them. Audit them. That still works. It is the safest option for a fleet. Lend covers the other case. One developer wants one agent on one remote host to act as them. They do not want to copy a key the agent can read at any moment. The approval prompt is the trade. A long run with many small writes becomes a long run with many small taps.

Source

Antonio Mika, "Running coding agents without handing them my credentials". Hacker News discussion, 82 points at the time of writing.

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project