lend Proxy Lets Coding Agents Reach GitHub Without Handing Them Your Keys: Everything Goes Through Touch ID
Coding agents on disposable cloud VMs need tokens and SSH keys, but storing them there is risky: any process on the machine can read them. The lend proxy keeps the secrets on your laptop. When an agent asks to run a command such as gh pr view, the command executes on the laptop after you confirm with a fingerprint (Touch ID), and only the output goes back to the agent. So instead of copying credentials to the VM, you route requests through an allowlist of permitted commands.
AI-processed from Hacker News AI; edited by Hamidun News
Imagine you've launched several coding agents, and each one needs to hit GitHub, call an API with a token, or SSH into another server. The agents run on disposable virtual machines (VMs) in the cloud, while all your credentials sit on your laptop. How do you let an agent run something like gh pr view without giving it the token itself? Developer antoniomika wrote the lend proxy to do exactly that. The agent asks to run a command, the command runs on your laptop, and you confirm it with your fingerprint (Touch ID on a Mac). The agent receives only the result.
Why the usual approaches don't work
The author runs Claude Code, Codex, and background jobs on virtual machines he wouldn't mind throwing away. The laptop keeps the SSH keys, the 1Password vault, and the cloud CLIs he's logged into. While the agent is editing code, everything is fine. Problems start when it needs access on your behalf.
There are usually two paths, and the author dislikes both. The first: copy the token to the VM. Then any process on the machine can read it, including the agent itself and everything it launches, for as long as the token is valid. The second: forward the SSH agent with ForwardAgent. Private keys stay on the laptop, but the VM can use every key without asking and without a log of signatures. Moreover, this only works for SSH keys, not for API tokens.
What the author needed was different: credentials stay on the laptop, and the commands that need them also run there. The agent asks, a human confirms, and the agent gets the result.
How lend works
lend is a proxy in front of a regular SSH agent. You forward its socket to the VM instead of the real one, so SSH and Git work as before. In addition, the proxy can execute commands through an extension of the SSH agent protocol. When an agent on the VM calls lend exec -- gh ..., the request travels over the socket, and gh runs on the laptop under your login. The token never leaves, and stdout, stderr, and the exit code are returned to the agent.
Every request, whether a key signature or a command, is checked against the policy on the laptop. The policy lists what the VM may ask for at all. Everything else is denied. You can permit not the whole program but individual subcommands, for example only gh pr view. The approval dialog shows the full path to the executable, all arguments, the working directory, the reason, and what the process reported about itself. The reason and caller data are sent by the VM itself, so they are marked as unverified. They help make a decision but do not change what the policy allows.
For Claude Code there is an MCP server with a single tool, which requires a command and a reason. Without MCP, you can call the regular CLI.
Secrets can also be injected without the agent's involvement. A command rule reads a value, for example from 1Password, and passes it to the approved command as an environment variable. The token stays on the laptop. The VM chooses neither the variable names nor the secret references. Values do not appear in dialogs or history. There is also a policy for keys: unlisted keys are invisible to the VM, and for listed keys you can limit what they sign and how many times per session.
There are temporary grants: after approval you can allow, say, five minutes or three runs, but only for exactly the same command with the same arguments. They disappear when the session ends or the screen locks. A separate command shows what is connected and what is awaiting a response, and an emergency stop cancels everything and rejects new requests. The audit log by default contains no arguments, output, or secrets.
One more practical detail: an agent's questions such as "may I run this command on the VM?" can also be sent to the same dialog on the Mac via the lend approve hook. There are adapters for Claude Code, Codex, Antigravity, and pi. If the laptop is unavailable, the hook does not get in the way, and the agent asks in its own interface.
Limitations
The author openly describes the weak spots. lend does not isolate execution: an approved command runs as your user and sees your files. If you approve a shell, an interpreter, or a program that reads stdin, effectively everything is allowed, because stdin arrives after confirmation and is not visible in the dialog. If a command prints its environment, the secrets will reach the agent. That is why you must not allow env or op read as standalone commands.
Anyone with access to the forwarded socket on the VM, including root, can send requests while the session is open. The policy and confirmation still apply in that case. If the SSH connection drops, requests fail, and after reconnecting a new session starts with new approvals. The project is early, runs on macOS and Linux, and does not allocate a PTY, so interactive programs won't work. Biometrics are available only on macOS; on Linux, confirmation goes through Zenity or the terminal.
For teams that launch agents on remote machines, the takeaway is simple: don't copy tokens to VMs, and don't forward your entire SSH agent without controls. It's better to keep credentials where you are logged in and pass each request through a narrow allowlist of permitted commands. And keep in mind that an approved command runs with your privileges.
- Nov 21, 2026Anthropic Raises $81B in IPO at $2 Trillion Valuation as Protesters Show Up Outside Its Offices
- Nov 21, 2026Banks Stand to Earn $550 Million on Anthropic's IPO, More Than on SpaceX
- Nov 21, 2026Anthropic's IPO Order Book Is Several Times Larger Than the Shares on Offer; Founders Set to Receive $1.75 Billion
- Nov 21, 2026Anthropic Researcher Claims He Proved the Riemann Hypothesis with Claude in Six Weeks
Key Questions Answered
Why the usual approaches don't work?
The author runs Claude Code, Codex, and background jobs on virtual machines he wouldn't mind throwing away. The laptop keeps the SSH keys, the 1Password vault, and the cloud CLIs he's logged into. While the agent is editing code, everything is fine. Problems start when it needs access on your behalf. There are usually two paths, and the author dislikes both. The first: copy the token to the VM.
How lend works?
lend is a proxy in front of a regular SSH agent. You forward its socket to the VM instead of the real one, so SSH and Git work as before. In addition, the proxy can execute commands through an extension of the SSH agent protocol. When an agent on the VM calls lend exec -- gh ..., the request travels over the socket, and gh runs on the laptop under your login. The token never leaves, and stdout, stderr, and the exit code are returned to the agent.
What are the key details on Limitations?
The author openly describes the weak spots. lend does not isolate execution: an approved command runs as your user and sees your files. If you approve a shell, an interpreter, or a program that reads stdin, effectively everything is allowed, because stdin arrives after confirmation and is not visible in the dialog. If a command prints its environment, the secrets will reach the agent. That is why you must not allow env or op read as standalone commands.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.