How to let Claude Code manage your servers — without giving it your SSH keys

AI coding tools are stunningly good at diagnosing server problems — and completely blocked the moment the answer lives on a real server. The usual workaround is ugly: paste your terminal output into the chat, copy the suggested command back, repeat.

The obvious fix — giving the AI your SSH credentials — is a terrible idea. Here’s the version that isn’t.

How it works

Sshly runs a small local MCP server on your Mac. An MCP-capable tool like Claude Code connects to it and sees your servers by name only:

1. The AI asks Sshly to run a command on, say, web-prod-02.
2. Sshly asks you first — an approval prompt shows exactly what will run, where.
3. On approval, Sshly opens the connection with the credentials from your Keychain and returns only the command output.

Your password or key never enters the AI’s context. It can’t be logged, leaked or memorized — the AI literally never sees it.

Setup (about five minutes)

1. Open Sshly → Settings → AI Access and enable the bridge.
2. Copy the ready-made setup command shown there and run it in your terminal — it registers Sshly as an MCP server in Claude Code.
3. Ask Claude something real: “Why is web-prod-02 slow?”

Claude will request a command, you approve it, and the loop that used to be copy-paste becomes a conversation.

What this is good for

• Debugging: “Check the nginx error log for anything unusual in the last hour.”
• Ops chores: “Renew the cert and reload nginx.”
• Audits: “Which servers still run Ubuntu 20.04?” — across every connection you approve.

Every command is individually approved, so the AI has exactly as much power as you hand it, one command at a time.

Get Sshly → — the free tier includes the full MCP bridge.