Aegitox

Discord’s native AutoMod operates on pattern matching: if the token does not match the list, the message passes. Aegitox replaces that model entirely. When a message is sent, the NativeIntentEngine intercepts the payload at the Discord gateway layer, runs it through a primary MiniLM-L6-v2 model for toxicity detection, and — if flagged — hands off to a secondary model that classifies the threat vector as Personal, Self, or System (0-indexed). The original message is never delivered; instead, one of 120 target-specific placeholder responses is substituted in constant O(1) time. The whole cycle, per the vendor, runs in 2–12 milliseconds using a local edge architecture that does not touch a cloud API during inference.

The differentiating mechanism is what the vendor calls the Head + Tail token splice: for longer messages, the pipeline extracts the first 32 and last 32 tokens, runs intent mapping on that structure, and classifies intent mathematically rather than lexically. This is what allows it to catch unicode evasion tactics — zero-width joiners, hidden markdown, and similar obfuscation that breaks static regex — because the model is evaluating conversational intent, not surface-level character sequences. Webhook token-stealer detection is handled as a separate module, intercepting malicious payloads before execution.

The tool fits large Discord guilds that need autonomous, sub-second moderation without a team of moderators reviewing every flag in real time. The karma-based automated penalty engine — which triggers timeouts based on configurable thresholds and routes appeals directly to staff via DM — means a human still reviews before a penalty becomes permanent. That architecture works well for communities where procedural fairness matters. It breaks down for teams that need self-hosted deployment, external API access to moderation decisions, or integration with their own analytics stack — none of those are available, and the vendor page describes no roadmap for them.