AI features

Providers, data controls, safety and the Server Builder.

The operator chooses a text provider (OpenAI-compatible endpoint, Anthropic, or Ollama for local models) and optionally an image classifier. Each server decides whether message content may be shared with the provider; without that permission nothing is sent.

Content is treated as untrusted data: it is wrapped in randomly delimited blocks, the model is told never to follow instructions inside it, it has no tools, and its output must validate against a strict schema. Deterministic code decides the outcome: at most flagging a message for review or deleting that one message when you enable it. AI never bans, kicks or changes configuration. Budgets, timeouts and a circuit breaker keep cost and latency bounded, and provider status is visible in the dashboard.

The Server Builder generates three proposals from your description, validates them (no Administrator or server-management permissions, bounded sizes, consistent references), previews them, and applies the one you approve through audited bot jobs that record every created resource for rollback.