Safety and Control

The Short Version

The DevOps AI Agent works on your server with full root access, and it is an AI: it can be wrong. These habits keep you in control.

• Take a snapshot before a bigger change. Restoring it undoes everything written since.
• Press Stop whenever you want the agent to stop.
• Read the tool rows. They show every command it ran and every file it wrote.
• Keep secrets out of the conversation unless the agent asks for one, and use keys you can revoke.
• Know your own way into the server: the Browser Console, and Reset for the root password.

Snapshots

A snapshot is a copy of the server’s disk, so restoring one rolls back everything on it: the agent’s work and your data alike. Press Create Snapshot on the server’s manage page before upgrades, configuration changes, and experiments. You can keep up to 5 snapshots, at $0.10 per GB per month. A snapshot is kept with the server, so it is not an off-server backup. See Snapshots and Restoring a Snapshot.

Snapshots are yours to manage: the agent has no tool to create, delete, or restore one.

Stopping the Agent

While the agent is working, the Send button becomes Stop. Stop ends the run and closes the agent’s connection to your server. A command it had started may or may not have finished, so ask the agent to check the state of the server before it goes on. You can write to it again whenever you like.

What the Agent Can Reach

The agent is given the login of this one server and tools to work on it over SSH: running commands, writing files, and reading files. Everything it does is something a person logged in as root could do there, which includes deleting data.

It has no tool for your RepoCloud account, your balance, your payment cards, your DNS provider, or your other servers. It cannot change the tier, create a snapshot, or press any button on the manage page.

Who Sees What

The agent is an AI language model. To answer, it is sent your messages, its own earlier replies, and the output of every command it runs and every file it reads. That output becomes part of the conversation, so asking the agent to print a file that holds secrets, such as an environment file, puts those secrets into the conversation. RepoCloud chooses the model.

The conversation is saved, which is how Resume Chat shows it again. For how RepoCloud handles your data, read the Privacy Policy.

Keeping Your Own Way In

The agent logs in to your server over SSH as root, with the root password RepoCloud holds. A change that closes that route also closes the agent’s route: a new root password, SSH login settings that turn off the root login or the password login, a different SSH port, or a firewall rule that blocks SSH. The server and its application keep running, but neither you nor the agent can log in the usual way.

You can keep that from happening and you can recover from it:

• Tell the agent in advance. Put “Do not change the root password, the SSH settings, or the SSH port” in the Custom Instructions, or write it in the chat before a change that touches the network.
• Use the Browser Console. Open Console on the manage page opens a browser terminal on the server’s serial console. A serial console is attached to the server’s serial port and not to its network, so a firewall rule inside the server does not apply to it. It asks for the root password, so reset the password first if it was changed inside the server. See Browser Console.
• Reset the root password. A password changed inside the server is not known to RepoCloud, so the one on file stops working. Reset on the manage page sets a new root password, emails it to you, and stops the old one. After that, press Resume Chat. See Root Password.

Good Habits

• Ask for the plan first when a change is big: “Tell me what you will change, and wait for my go-ahead.”
• Say what must not be touched: a database, a folder of uploads, a configuration you edited yourself.
• Check the result yourself. The agent checks logs and HTTP responses, and it cannot see the page as you see it.
• Give keys the least permission they need, and revoke the ones you no longer use. A key you gave the agent may now be written in a configuration file on the server.
• Tell the agent what you changed by hand. It knows only what it did and read in the conversation.

Next Steps

• The DevOps AI Agent

  What the agent does and how the chat works

• Customizing Applications

  Recipes for changing an application, with snapshots first

• Troubleshooting the Chat

  When the chat does not open, the agent is stuck, or a server will not answer

• Managing VPS Apps

  Snapshots, console, root password, and the rest of the manage page