Lend Gates Remote Coding Agents Behind Local Approval
Software / news
Lend Gates Remote Coding Agents Behind Local Approval
Antonio Mika's MIT-licensed proxy keeps SSH keys and secrets on the laptop, but the README says an approved command is not sandboxed.

Developer Antonio Mika released lend, a proxy that sits in front of an SSH agent so that a coding agent on a remote machine can run commands on your laptop only after a local policy, and sometimes a Touch ID prompt, approves each one.
The tool is described in a blog post dated Nov. 13 and updated Nov. 17, and the code is in a GitHub repository under the MIT licence. It is early: the repository showed 11 stars, no forks and an empty Releases section on Nov. 17, and the README calls it an initial implementation for macOS and Linux that needs Go 1.24 or later.
How lend sits between the agent and the keys
You forward lend's socket to the remote host instead of the real SSH agent socket. SSH and Git keep working, and lend adds an agent extension for running commands. Every request, a signature or a command, is checked against a policy file on the laptop, and anything not allowed is denied.
The example policy in the repository shows the shape. The top level is "default": "deny" with "require_reason": true. One rule lets /opt/homebrew/bin/op run with the prefix read behind a biometric prompt, and another lets /opt/homebrew/bin/gh run pr view behind a confirmation. The ssh section is also deny by default, with one key rule capped at 10 signatures.

| Command | What it does |
|---|---|
lend ssh | Starts a proxy for one SSH connection |
lend exec | Runs an approved command from the remote host |
lend mcp | MCP server exposing a run_local_command tool |
lend approve | Routes an agent's permission prompts to the laptop |
Secrets that never reach the remote machine
Mika's pitch is that the token stays on the laptop. A configured secret reader, such as op read, fetches a value and injects it as an environment variable into an approved command. The remote side cannot choose which variables are set or which references are read, and the approval dialog shows variable names and references, not values. In the post's words, "The token never leaves the laptop, so the agent never has it."
The case for that design is visible in two incidents this site has covered. The Fireworks AI breach reached Uber, Cursor and Notion, and Mandiant's LLMjacking findings describe stolen cloud keys. Lend does not address either incident directly. It addresses the narrower case of an agent holding a key it does not need to hold.
Where approval stops protecting you
The README and the post are blunt about limits, and one of them is easy to miss. An approved command runs as your user and is not sandboxed. Standard input streams after approval and is not shown in the prompt, so approving a shell or interpreter approves arbitrary code. The post warns against approving env, sh or op itself for that reason.
Caller details in prompts come from the remote VM and are unverified, and anyone who can reach the forwarded socket, root included, can make requests while the session is open. Interactive programs fail because lend allocates no PTY, and biometric approval works only on macOS.
lend approve extends this to the agent's own permission prompts. It takes Claude Code's PermissionRequest hook, which returns an allow or deny decision, and sends the question to the laptop. Adapters exist for Codex, Antigravity and pi. The post does not say what happens when the laptop is unreachable. Claude Code's hooks documentation says a command hook can set onFailure to "block" so failures deny the request, which requires Claude Code v2.1.295 or later.
Mika credits ssh-rev-exec and op-forward as prior art and says lend adds per-request policy, local approval and key filtering. The repository lists no tagged release yet, so the first version number is the thing to watch.
Sources
More in Software
- 01Zigzag, a 35 MB Mac Browser, Courts Arc Users 35 Days Before ShutdownThe WebKit-based browser imports Arc spaces and pinned tabs, but its README lists no sync, no Boosts and a partial Chrome extension API.
- 02Nimony 0.6.2 Sets a Year-End Goal for Nim 3 With 44 Days LeftThe compiler that will become Nim 3 is MIT-licensed and builds natively without LLVM, but its FAQ says it cannot yet compile Nim 2 code unchanged.
- 03Alloy CSS Runs Menus and Tabs Without JavaScript, but Dialogs Need Safari 26.2The 24 KB MIT stylesheet leans on popover and invoker commands, so its no-script promise has browser floors that differ by component.
- 04Windows 11 Telemetry Flags Which Coding Agents You Run, Researcher FindsXusheng Li traced a service-experiment feature in build 26200.9974 that matches executable names such as claude.exe and queues the result as optional diagnostic data.