AegisAI
Kotbo × AegisAI has every message scored by AegisAI, the AI built by marvideo. A toxicity score from 0 to 100 decides what happens next: nothing, a check by staff, or immediate action by the bot. Emotion analysis also feeds conflict and distress alerts, and the server’s tone in the statistics.
Turning analysis on
Section titled “Turning analysis on”The Message analysis switch at the top of the page turns the module on. Its status shows next to it: whether AegisAI is reachable, its response time, and the queue size when messages are waiting.
The first time you turn it on, Kotbo asks Help train AegisAI?:
- Yes, share: analyzed messages join AegisAI’s training set, to improve its detection;
- No, keep nothing: they are scored, then forgotten.
You can change that choice later in the Privacy block. Until it is made, nothing is shared.
The page has three tabs: Settings, To review and Modules.
Settings
Section titled “Settings”Sensitivity
Section titled “Sensitivity”Two thresholds split the toxicity scale:
- below the Review threshold, nothing happens;
- between the two, the message is sent to you under To review;
- above the Automatic action threshold, the bot acts on its own.
The Preset sets both thresholds at once:
| Preset | Review threshold | Automatic action |
|---|---|---|
| Lenient | 88 | 97 |
| Balanced | 80 | 95 |
| Strict | 70 | 90 |
Changing a threshold by hand switches the preset to Custom. The second threshold must stay above the first.
The Try a sentence field scores a message of your choice against the thresholds shown: Analyze returns its Toxicity and its Emotion. The sentence is only shared for training if you agreed to it.
When the bot acts on its own
Section titled “When the bot acts on its own”What happens above the automatic action threshold:
- Action: Remove the message, Remove and warn or Remove and time out. The warning has a Warning weight (Light, Normal, Serious) on the weighted warning scale; the timeout, a length in minutes.
- Notify the member: a short message in the channel, deleted after a few seconds.
- Card channel: where messages to review and the bot’s actions are posted. Empty, it is the server’s log channel.
A message scored late goes to review instead of being removed. Every sanction issued by AegisAI attaches a transcript of the channel as evidence, taken before the message is removed.
What is analyzed
Section titled “What is analyzed”Every message is scored, staff and administrators included. Three options extend the analysis:
- Edited messages: catches insults added afterwards to get around the filter.
- Nicknames: a nickname judged toxic is replaced with a neutral one, and its owner gets the automatic action’s sanction.
- Ticket mood: shows the requester’s mood on each ticket. High priority on anger or distress raises a low or normal priority ticket to high.
Exemptions
Section titled “Exemptions”Exempt channels and Exempt roles: their messages are scored, but never removed or sanctioned automatically. Above the threshold, they go to review. Administrators are treated the same way.
Privacy
Section titled “Privacy”The block recaps what goes to AegisAI (the text, without mentions, links or emojis) and what Kotbo keeps: counters per hour and per channel, and the excerpt of flagged messages, erased when the log retention period ends. This is where you change Help train AegisAI.
To review
Section titled “To review”The tab lists detections, with a counter of pending messages. The Status filter shows To review, Automatic, Confirmed, False positives or All.
Each card shows the channel (or nickname, or edited message), the Toxicity score, the emotion, the person targeted if any, and the message excerpt. Scored late flags a message scored too late to be removed.
| Button | Effect |
|---|---|
| Confirm and sanction | Applies the planned sanction. For a conflict or distress alert, the button reads Handled. |
| False positive | Undoes what the bot did. For a distress alert: False alarm. |
| Lift slowmode | Removes the slowmode set by an escalation. |
| View message / Evidence | Open the message on Discord and the attached transcript. |
The same cards are posted to the Card channel on Discord, with the same buttons (labelled in French): staff can rule without opening the dashboard.
Modules
Section titled “Modules”Three detections build on the scores. Each has its own switch.
Escalation
Section titled “Escalation”When things get heated between several members of a channel, the bot sets a temporary slowmode and alerts staff. Settings: Heated messages to trigger, Window (seconds), Slowmode (seconds) and Slowmode duration (minutes).
Targeted harassment
Section titled “Targeted harassment”A member who sends several toxic messages to the same person, as replies or mentions, is flagged along with their messages. Settings: Toxic messages to trigger, Window (minutes), and the Action - Alert staff or Time out and alert.
Two more thresholds watch over time, all authors combined: they spot a group ganging up on one person, or attacks spread over several days. They alert staff, with no sanction.
Distress
Section titled “Distress”A message with strong sadness or fear is quietly flagged to staff. Never any sanction. Settings: Emotion threshold, Alert channel (a restricted channel ideally; empty, the card channel) and One alert per member every (hours).
Tone in the statistics
Section titled “Tone in the statistics”Detected emotions are only kept as aggregates, and only if the server’s statistics are turned on. They feed the Tone section of Statistics and the tone card on a member’s profile.