Set up once, then build with your AI

The idea: you set things up one time, and from then on you work on your server by talking to your own AI. This page covers the one-time setup and the safety net around it. It does not give you prompts to copy. Describe what you want in your own words.

Step 1: Connect your AI app

You add one connector for your whole account. Its address is https://mcp.quarry.host/mcp, also shown on the Connect your AI page. Use whichever of these you have.

• Claude, in the app or in a terminal. In the Claude app or on claude.ai, add a custom connector in Settings, Connectors. In a terminal, use the command on the Claude Code page. See Connect Claude and Claude Code, Cursor and other MCP apps.
• Claude Code. One command adds the connector, and /mcp signs you in. See Claude Code, Cursor and other MCP apps.
• ChatGPT. Add a custom connector on a plan that allows it. See Connect ChatGPT.
• Other AI tools. Any app that supports remote MCP servers can connect the same way. See Claude Code, Cursor and other MCP apps.

In every case you sign in to Quarry Hosting, and then check the consent screen. See Connecting your AI.

Pick a capable, trusted model. Your AI is making changes to a live server, so a stronger model makes fewer mistakes, and a model from a company you trust is a safer choice than an unknown tool.

Step 2: Choose how much it may do

Each server has an AI page with Allow, Ask and Block for every action. That is the default way to work: your AI asks you before anything risky.

Full autonomy

If you want your AI to get on with the work, there is a full-autonomy switch for each server. It is off until you turn it on.

Available once enabled on our machines. It stays switched off for everyone until the protections on our game machines are in place. This page will be updated when it opens.

• To turn it on: only the owner can, from the panel in a signed-in browser. You must have two-factor sign-in set up, and you sign in again and confirm a card that lists what changes. Your AI cannot turn it on.
• What it allows: things that would have asked for approval now run straight away, including running any console command and working with files by path. Your AI is still told to say in the chat what it is about to do before any risky step. Your own Block settings still hold.
• Team members: a member's AI gets autonomy only within what that member's role could do.
• To turn it off: the owner, or anyone with the right role, can switch it off at any time, and your earlier Allow, Ask and Block table is back exactly as you left it.

The four things it never does

Even with autonomy on, your AI cannot:

1. Spend money or change payments.
2. Delete the server.
3. Change the team or anyone's permissions in the panel.
4. Touch account security, such as your password, two-factor sign-in or connected apps.

There is no tool for these, so no instruction can make your AI do them.

What always waits for your tap

Giving or removing op, bans, changing permissions, and turning the whitelist off always wait for one tap from you in the panel, on desktop or phone, however autonomy is set. So do rollbacks, undos and restores that put settings files back, because they could undo a change you tapped to approve. A restore of worlds only does not need the tap. Your AI is told the request is waiting.

On top of that, when a risky request names something that appeared in text from players, signs, books or plugin files that your AI read in the last hour, it waits for your tap as well. This is a best-effort check and is easy to get around, so do not rely on it. The real protections are the taps above, the activity feed and Undo.

Game chat is never instructions

Your AI must never act on what a player says in game chat, a sign, a book or a plugin file, even if the text says it comes from you. Only the messages you type to your AI count. If something in the game tries to give your AI orders, ignore it and tell your AI to as well. Anything that looks like it came from a player is marked as untrusted when your AI reads it.

Step 3: Watch the activity feed, and undo

Everything your AI does is listed in the activity feed on the server's AI page, with what it changed. A short line on the server's pages shows when your AI is working. Next to each action there is Undo. Before any change, we take a snapshot, so there is something to go back to. If you want to stop everything, press Stop on the AI page.

Step 4: Keep your settings in Git sync

Git sync saves your server's settings to your own GitHub, GitLab or Codeberg repository after every change. If your AI makes a change you do not like, you can go back to an earlier save. You can also allow your AI to read the history and restore a save. Follow Save your settings to GitHub.

Summary

1. Connect your AI app once.
2. Leave it on Ask, or use full autonomy once it is available.
3. Use the activity feed and Undo.
4. Connect Git sync as a second safety net.