AI assistant
Integrations & actions
Connect your game server panel with allow, ask and deny per tool.
What integrations are#
Integrations connect Tikkora to your game server, server panel or plugins, so the assistant can look things up (“is this player online?”, “when did they last join?”) and staff can act straight from a ticket (whitelist, mute, ban). They use the open Model Context Protocol (MCP): any MCP server that speaks HTTP works, and each one offers a set of tools.
Connecting a server#
In Discord, open /config area: Integrations and press Add. Give it:
- A name, like “Server panel”.
- The URL of the MCP server, using Streamable HTTP or SSE.
- Optionally an auth header name (e.g.
Authorization) and secret. The secret is stored encrypted.
The bot lists the server’s tools so you can decide what each one may do. It checks every integration every 10 minutes; if one becomes unreachable, admins are told and its tools are unavailable until it recovers. /status shows whether every integration is reachable.
Allow, ask or deny per tool#
| Permission | What the AI may do |
|---|---|
| Allow | Call the tool on its own, e.g. look up a player. |
| Ask | Request it. A staff member must approve each call before it runs. |
| Deny | Nothing. The AI never sees the tool. |
Safe defaults
- New read-only tools (lookups like
get_,list_,search_) start as Allow. - New tools that change something start as Deny until you decide. A button applies our suggestions to everything still denied: Allow for reads, Ask for the rest.
- Dangerous tools can never be set to Allow, only Ask or Deny. That covers console commands, restarts, file changes, bans, kicks and mutes, permissions and ranks, and economy changes. A tool marked read-only by the server still counts as dangerous if its name says otherwise.
- If a server changes an allowed tool so it no longer looks read-only, it drops to Ask automatically.
You can also override the permission per category, e.g. allow a lookup in Bug reports but deny it in Appeals.
Who can approve#
Each tool has a lowest tier that may approve or run it. The bot suggests one from what the tool does, and you can change it:
| Suggested tier | For tools like |
|---|---|
| Management | Console and commands, restarts, files, permissions and ranks, economy. |
| Staff | Ban, unban, kick. |
| Support | Everything else: lookups, mute, warn, whitelist… |
Approving a request#
When the AI wants to use an Ask tool, an approval card appears for staff with the tool, the arguments and the AI’s reason, and the ticket shows “Waiting for staff to approve”. Staff can approve, deny, or edit the arguments and then approve. An approved call runs once, right away.
- You can’t approve actions on your own ticket.
- Requests nobody decides within 30 minutes expire, and the ticket is handed to staff.
Running actions yourself#
Staff can run any tool of a connected integration straight from a ticket with the Actions button on the staff notification, as long as their tier covers the tool. If the category has suggested actions on, the AI’s suggestions come with ready-to-run buttons. Every run is recorded on the ticket, so learned solutions remember what staff did and can suggest it next time.
Linking Minecraft accounts#
Some tools should only ever act on the player who opened the ticket, like “restore my inventory”. Open the tool and bind its player argument to the ticket owner. The AI can then only use it with the owner’s linked Minecraft name.
- Players link themselves with
/link me name:(3–16 letters, digits or _). - Staff can set and verify an account for someone with
/link set member: name:. /link showshows your linked account.
Safety limits#
- The AI may make at most 20 tool calls per ticket.
- Practice answers (learning mode) only use allowed read-only tools.
- The AI treats tool results as data and ignores any instructions inside them, and never says an action is done unless a result confirms it.
- Every call, approval and staff run is logged with who, what and the result.