AI Assistant
The AI assistant completes routine operations in natural language, such as querying players, adjusting health, running backups and reading logs. Every tool call is gated by permissions and switches, and written to the audit log.
Model configuration
Under AI Settings, provide the service endpoint, API key and model name. Services compatible with the OpenAI Chat Completions protocol are supported.
After saving, start a conversation. Connection errors are reported explicitly rather than failing silently.
Tool switches
Each tool definition includes:
- Tool ID and description.
- Required permission.
- Default switch state.
- Current availability and the reason when unavailable.
Tools enabled by default cover daily queries and player actions, including player lists, player actions, documentation search and the tool catalog. Tools that execute commands, write files or call arbitrary APIs are disabled by default and must be enabled manually.
When a tool is disabled, the assistant can still see its definition and purpose, but calls are rejected with an explicit reason. The assistant should direct the user to AI Settings → Tools instead of trying another route.
Tool catalog
Ask the assistant to list tools, or use the tool catalog tool to view:
- ID, description and required permission for every tool.
- Whether it is enabled and whether it is enabled by default.
- The reason when unavailable.
- Whether the current account holds the required permission.
Approval
High-risk calls can use manual or automatic approval:
- Manual: an approval card appears in the conversation before execution.
- Automatic: a model checks whether the call matches user intent and returns a reason for approval or rejection.
- Rejection reasons: every rejection includes a clear reason so the assistant can adjust.
Approval decisions are recorded in the audit log.
Permission model
Panel accounts have three roles:
| Role | Description |
|---|---|
| Super administrator | All features and panel settings |
| Administrator | Instance and player management within granted instances |
| User | Read-only or explicitly granted operations |
Instance permissions cover files, console, instance control and resources. Plugin endpoints have separate permissions: player view, player manage, inventory edit, world manage and console.
AI tool calls use the permissions of the conversation owner and never exceed that account.
Security notes
- The panel directory and listening ports are invisible and inaccessible to AI tools.
- Command tools run inside a restricted workspace.
- Every write action records operator, arguments and result.
- Enable high-risk tools only for trusted accounts.
Troubleshooting
The assistant says it lacks a capability
Check the tool catalog for switch state and permissions. Enable default-off tools manually, or ask an administrator to grant the missing permission.
Calls hang
Check the model endpoint, network connectivity and API key. Long-running tasks may time out; adjust the timeout in settings.
Automatic approval rejects a call
Read the rejection reason. If it does not match your intent, switch to manual approval or rephrase the request.