01
V3
Reasoning engine
Reads a message alongside the previous twenty, the relationship between accounts and the apparent intent.
V3 — new reasoning engine
Moderation that reads
the intent, not the wordlist.
ONIA V3 reads a message the way a human moderator does — the thread around it, the intent behind it, the history between the two accounts. Not a list of banned words.
Free to start 16 languages No personal data collected
Decision pipeline
01
Message received
websocket
02
Local prefilter
0.4 ms
03
Thread context
20 messages
04
V3 model
~150 ms
05
Action applied
idempotent
- 10M+
- messages protected
- 1,000+
- active servers
- 99.9%
- uptime
- 16
- languages understood
What changes in V3
Every system rebuilt, not repainted.
V3 is a full rewrite of the moderation core. Here is the honest before / after, system by system.
System
V2
V3
Analysis engine
per-message classification
reasons over thread, intent and history
False positives
jokes and irony often punished
tone and member relationship taken into account
Anti-raid
fixed thresholds, manual response
automatic lockdown in 2 s + quarantine
Warnings
plain counter
tiers, durations, expiry and full history
Attachments
extension filter
extension + binary signature + heuristics
Configuration
step-by-step /config menu
single panel, applied without restart
Logs
action logged
action + model reasoning logged
Languages
per-server detection
per-message detection, 16 languages
ONIA in action
Five scenarios, exactly as they run.
Not a marketing render: the same decisions, timings and reasoning your server gets. Pick a chapter.
ONIA in action
0:00 / 0:12
A message lands in the channel.
CHAPTER 01 / 05Systems
Eight modules. One configuration panel.
02
CORE
Moderation levels
Lenient, moderate or strict. Same engine, different tolerance for your server culture.
03
PREMIUM
Custom prompt
Describe your rules in plain language. The model applies them — no regex, no wordlist.
04
CORE
Tiered warnings
Each tier triggers the action you choose: warning, timeout, kick or ban.
05
PREMIUM
Anti-raid
Mass joins, spam, role and channel abuse, automated accounts, admin action abuse.
06
PREMIUM
Attachment scanning
Three layers: allowed extension, real binary signature, heuristics on name and structure.
07
CORE
Exemptions
Roles and channels excluded from analysis, so staff and private channels stay untouched.
08
CORE
Explainable logs
Every decision is logged with the reason behind it. Your moderators can contest it.
Infrastructure
Boring where it counts.
A moderation bot is infrastructure. It has to be up during the raid, silent the rest of the time, and it must never become the incident itself.
- Measured uptime
- 99.9%
- Decision latency (median)
- 190 ms
- Messages protected
- 10M+
- Languages understood
- 16
- Failover
- automatic
Data we never store
- Email address
- IP address
- Location data
- Usernames of sanctioned members
- Messages that were not moderated
What we keep
- Server ID and configuration
- IDs of sanctioned accounts
- Sanction history (type, date, reason)
Commands
Everything runs from Discord.
/config
ADMIN
Single panel: AI level, logs, tiers, exemptions, premium.
/antiraid
ADMIN · PREMIUM
Status, activation, allowlist and manual unlock.
/warnlist
MOD
Detailed history, pagination and bulk removal.
/ping
EVERYONE
Discord REST, WebSocket and AI model latency.
Two minutes to install. It learns your server, not a wordlist.
Add ONIA, run /config, pick a level. You can undo everything with one command.