Save Your SSH Connections,
Run AI Agents on the Remote Box

How can my AI agents access my SSH connections?

Through AgentsRoom MCP. Your saved connections are exposed to every agent as five MCP tools : ssh_list returns the servers usable in the current project (SSH, AWS SSM and WinRM alike), ssh_exec runs one command on a server and hands back its output (a PowerShell line on a WinRM host, read-only unless you allow writes, which you confirm one by one, only when risky, or never, as you choose), ssh_transfer moves one file to or from a server, ssh_connect opens one as a real terminal tab, and ssh_connection_new proposes a server you have not saved yet. So when you tell an agent "connect to prod and read the logs" or "list the Exchange send connectors", it does not have to guess a hostname or write a command by hand : it reads your connection list and works on the right machine.

Agent SSH access is deliberately credential-free. ssh_list only ever returns metadata: name, host, port, user, auth type and scope. Passwords and key passphrases stay encrypted in your OS keychain, private keys stay on your disk, and the desktop app injects the secret into the terminal itself. The agent names a connection, it never handles it.

A session an agent opens with ssh_connect is a visible terminal tab you can read, type into and close at any moment, and a command it runs with ssh_exec comes back in its conversation with its exit code, so you always see what was run and what the machine answered. Files travel the same way : with ssh_transfer an agent pulls a dump or a log off a server, or pushes a file to it, through the same credential-free channel, and gets back a path on your machine rather than the file pasted into its conversation. A connection can be saved for a single project or globally, so an agent only ever sees the servers that make sense for what you are working on. Network appliances whose command line only works inside an open session, such as a WatchGuard firewall or many switches and routers, are covered too: ssh_exec opens the session, waits for the device prompt and types the commands one after the other.

While an agent works on a server, you watch it live. Every command it runs with ssh_exec scrolls in the Activity tab of the right panel, output included, in one journal per server; the agent's avatar shows on the Servers button and, with an Agent at work badge, in your Servers list. A command sent encoded is shown decoded. You can also follow it from your phone, in the agent's card of the mobile app. It only travels between your computer and your phone, end-to-end encrypted with secrets masked: nothing of what runs on your machines is stored on a server.

Creating can also be entirely the agent's job. ssh_connection_new keeps the reviewed path: it opens the creation form pre-filled with what the agent found in a compose file, a deploy script or your documentation, and you are the one who reviews it, types the credential and saves. Unless you turn off "Agents can manage connections" (Settings, on by default), ssh_connection_save writes the connection outright, and even then the credential never travels through the conversation: the agent names a secret from your vault and the desktop copies it. The off switch is written into the setting rather than asked for in a prompt, because a connection an agent could register unseen would be a backdoor.