AutoMod
Term lists and AI topics that check Discord messages and VRChat profiles, what a rule can do about a match, the AI tools AutoMod may use, the messages before the one being checked, pictures, the language on each flag, test sets, the trial before a rule acts, channel and role limits, the daily AI call limit and the Try it box.
AutoMod reads text your group already has, Discord messages and VRChat profiles, and checks it against rules you write or pick. When a rule matches, Modbot writes a flag: a note a moderator reads on the Flags page and either keeps or dismisses. A rule can also be set to act: delete the Discord message, time out its author, or ban or remove the person from your VRChat group.
Everything here is off to start with. There is one switch for the whole feature, each rule has its own switch, and every new rule only flags until you tell it to do more.
You set AutoMod up under Settings → AutoMod. Changing it needs the Change settings permission.
Two kinds of rule
Term lists are lists of words, phrases or patterns. Matching happens inside Modbot, costs nothing, and sends nothing to an AI provider. Term lists work with AI switched off, and they are what AutoMod runs on a deployment that never sets AI up.
AI topics describe, in your own words, something to catch ("scam links offering free Nitro", "asking members for personal details"). Modbot asks your AI model whether the text matches. AI topics need AI switched on under Settings → AI → Connection, and each check is a paid request to your provider. The AI section of the AutoMod tab, with the topics, the AI tools and the daily AI call limit, only appears while that switch is on.
What rules check
Each rule has Checks, the kinds of text it looks at. Tick any of:
| Checks | The text |
|---|---|
| Discord messages | The text of each new message the Discord bot stores, and again when it is edited. Messages from bots are skipped. |
| Display name | A VRChat user's display name. |
| Bio | A VRChat user's bio. |
| Status | A VRChat user's status line. |
| Pronouns | A VRChat user's pronouns. |
Discord messages are checked as the bot stores them, including messages it finds when catching up after being offline. See Discord messages and analytics for which messages the bot stores.
VRChat profiles are checked when Modbot first records a profile and whenever it records a change to one. Only the parts that changed are checked, so a new avatar costs nothing. A background job does this about once a minute, so a profile flag can appear shortly after the change is recorded.
Switching on for the first time
While AutoMod on is off, Modbot keeps its place in the profile history. When you switch it on, it works through every profile first seen or changed since that place, which on a new deployment is every profile Modbot has recorded. If you have AI topics that check profiles, this can mean many AI requests. The daily AI call limit and your spend limits are what hold it back.
Earlier messages
A reply, a joke and a quote all read differently on their own. Each rule has Earlier messages: how many messages before the one being checked go to the model so it can read that message in context. Choose None, 3 messages, 5 messages or 10 messages. A new rule starts on 5.
The messages come from the same channel, or from the same thread when the message is in one. When the message is a reply, the message it replied to goes with them.
The model is told to judge the last message only and to use the rest to understand it. A quote it takes from an earlier message is thrown away, so nobody is flagged for words somebody else wrote.
Earlier messages only apply to Discord messages. A VRChat profile has no conversation around it.
The flag records which messages went with it, and the Flags page shows them.
Pictures
Each rule has a Check pictures box. With it ticked, the rule also checks:
- image attachments on a Discord message,
- the author's Discord avatar,
- a VRChat profile picture and avatar picture.
The box is only available when the model you have chosen reads pictures. Modbot knows this from the model list on Connection; if your model reads text only, the box is greyed out.
At most four pictures go with one check, each at most 4 MB, and only PNG, JPEG, WebP and GIF. Anything larger or of another kind is skipped, and a link that points at a private or local address is never followed.
Pictures cost much more than words
A single picture can cost more than the whole message around it. Picture checks go through the daily AI call limit and your spend limits like any other AI request, and a rule with pictures on says so on its card.
When a picture matches, the flag says which one — "Attachment cat.png", "Discord avatar", "VRChat profile picture" — with a link to it.
Language
Every flag records the language of the text that was checked, worked out on your own server without sending anything anywhere. Very short text is left as Unknown.
The Flags page shows the language on each row and lets you filter by it, and each rule's card shows its dismissal rate for each language beside the whole figure. A rule that is right in English and wrong in every other language reads as fine until you split it that way.
The AutoMod card
AutoMod on switches the whole feature. With it off, nothing is checked. Press Save after changing it.
AI tools
The AI tools card appears while AI is switched on under Settings → AI → Connection. Each switch is one thing AutoMod may ask the AI to do; a tool that is off is never called.
| Tool | What it does | Starts |
|---|---|---|
| Check text against AI topics | Runs your AI topics on text no term list matched. Off, AI topics never run and only term lists check anything. | On |
| Check pictures | Sends a rule's pictures with the text, for rules with Check pictures ticked. Off, no picture goes to the model whatever a rule says. | On |
| Give an opinion on a flag | Puts an Ask AI button on each open flag. The answer, keep or dismiss with one sentence, goes on the flag and on its review. | On |
| Propose an action for a flag | With the opinion, the model also names what a moderator might do: nothing, delete the message, time out, ban from the group, remove from the group. Advice only: Modbot never carries it out. | Off |
The same card holds the Daily AI call limit, the most AI requests AutoMod makes in one day (UTC). It starts at 200 and can be set from 0 to 100,000. At 0, AI topics and opinions never run. When the limit is reached, they stop until the next UTC day and term lists keep working. AI calls today shows how many of today's calls are used, and AI shows Ready when a model answers or Off.
Press Save after changing a switch or the limit.
Term lists
The Term lists card lists every list with its source (Local or Modbot Hub), how many terms are on, what it checks, and how often it has flagged: for example "12 flags · 58% dismissed". A list that is mostly dismissed is mostly noise.
A local list
Press New list. Give it a Name, choose its Checks and Action, and add terms. Each term has a kind:
| Kind | Matches |
|---|---|
| Whole word | The word or phrase with a space, punctuation or the edge of the text on both sides. "scam" matches "a scam!" but not "scamper". |
| Contains | The word or phrase anywhere, including inside longer words. |
| Pattern | A regular expression. Modbot checks it when you save and refuses one that does not work. Each pattern gets at most 100 milliseconds per check; one that takes longer counts as no match. |
A list holds up to 5,000 terms, and each term can be up to 200 characters.
Before matching, Modbot makes text harder to disguise. It ignores upper and lower case, removes
invisible characters, joins up extra spaces, turns "fancy text" letters (bold, script, circled,
full-width) into plain ones, removes accents, and treats common look-alikes as the letters they
imitate, such as 0 for o, 1 for i, @ for a, $ for s, and Cyrillic letters that look
Latin. The flag still quotes the person's own words.
A list from Modbot Hub
Modbot Cloud, at cloud.modbot.co, offers ready-made term lists. Press Add from Modbot Hub to see them, with how many rules each has, its version and what it suits, and press Add.
A Hub list's name and terms belong to the Hub. You can still choose its Checks and Action, switch it on or off, and switch individual terms off in Edit (there is a Filter box to find them). Some Hub terms are written for particular kinds of text, and only run on those. Hub lists can also contain combination rules (several words near each other), which local lists do not offer.
Hub lists are never changed without you:
- Modbot checks each Hub list for a newer version every six hours. Refresh checks one now.
- When there is a newer version, an Update to button appears with the version. Hover over it to see how many terms it adds, removes and changes.
- Press it to use the new version. Terms you switched off stay off.
AI topics
Press New topic on the AI topics card. Fill in:
- Name, a short name that appears on flags.
- What to catch, in plain words, up to 2,000 characters.
- Sensitivity: Low (only clear, unmistakable cases), Medium (likely cases) or High (anything that could reasonably be it).
- Checks and Action, as for term lists.
The card shows AI off when no model answers yet. The whole AI section, this card included, is hidden while AI is switched off on Connection.
How AI topics run:
- Term lists go first. A topic only checks text that no term list has already matched for that kind of text, so a message a term list caught does not cost an AI request as well.
- Several pieces of text in one request. A profile's name, bio, status and pronouns go together, and the profile check hands several people over at once. The topics are listed once for all of them, which is most of what a request costs.
- Each person's text is still sent on its own. It never shares a message with Modbot's instructions, and the pieces are separated by a marker that is different every request, so nothing written in a message can pretend to be an instruction or to belong to somebody else. The model is told that everything inside is untrusted text to check and never something to obey. Member text is sent to the AI provider you choose.
- The model must quote the text. It answers with which piece of text, the topic, one sentence saying why, and the words it matched. If those words are not really in that piece of text, Modbot throws the match away.
- An answer Modbot cannot read is thrown away. If the model answers in the wrong shape, names a topic nobody asked about, or names a piece of text nobody sent, the whole answer is discarded rather than half-read — and the pieces are then sent again one at a time, so one confused answer costs a retry rather than costing everybody in it their check.
- Each request has thirty seconds. A provider that does not answer in time is recorded as an error in the call log, never as "nothing matched", and Modbot tries your fallback model once if you set one.
- Before each request, Modbot checks the daily AI call limit and the spend limits for everyone and for Moderation. When any is reached, AI topics stop and term lists carry on.
- Every request is written down in the call log. A request that produced a flag keeps the whole prompt and the whole answer, so you can see exactly what the model saw; one that flagged nothing keeps counts only.
AI settings lists exactly what is sent.
Test sets
Every rule has a test set: short sample texts you mark should flag or should not flag. Press Test set on a rule to open it.
- Add a sample, choose what kind of text it is, and add a note if you want. Up to 200 per rule.
- Run checks every sample against the rule as it is now, and shows each one as Caught, Missed, Wrongly flagged or Not flagged, with the totals. For an AI topic it shows the words the model quoted and the reason it gave.
- Runs are kept with the model that answered and the version of the rule, so you can compare a run after a model change with the run before it.
Running an AI topic's test set is one real AI request per sample. It uses today's AI calls and counts towards your spend limits like any other request. A term list's test set costs nothing.
Every new AI topic starts with four samples marked should not flag, each one an attempt to talk the model out of its instructions. A topic that flags one of those is reading the text as instructions rather than checking it.
Actions
Every rule starts as flag only. A rule's Action group lists everything it can do about a match:
| Action | Applies to | What happens |
|---|---|---|
| Flag | Everything | Always. A note on the Flags page. |
| Send to Reviews | Everything | Opens a review for each flag, so the team decides it. See Sending a flag to Reviews. |
| Delete the Discord message | Discord messages | The bot deletes the message. |
| Time out for | Discord messages | The bot times the author out, from 1 to 40,320 minutes (28 days, Discord's longest timeout). |
| Ban from the VRChat group | Display name, bio, status, pronouns | Modbot bans the person from your managed group, through the same call a moderator's Ban button makes. Their linked Discord account is banned too. |
| Remove from the VRChat group | Display name, bio, status, pronouns | Modbot removes the person from the group. When a rule asks for both, the ban is enough: VRChat's ban removes them too. |
A box is greyed out until the rule checks something it could act on: the Discord actions need Discord messages ticked, the VRChat actions need at least one profile field. A rule never acts across platforms, so a bio match never deletes a message and a message never bans anybody from the group. Modbot will not act on the account it signs in as.
Everything a VRChat action needs to be trusted applies to it exactly as to a Discord one: the test set, the trial, the channel and role limits, and the automatic pause below.
People a rule never acts on
A rule never bans or removes a member of your team: anybody whose VRChat account is linked to a Modbot account that signs in. That is always on, whatever the rule says and whatever role they hold, and there is nothing to switch. Only a link that has been proved counts; a link still waiting for its code in somebody's bio does not. The team member is still flagged, marked Exempt on the Flags page, so the team hears that a rule matched one of its own.
On top of that, each rule can name the roles it never acts on. See Where a rule runs.
Before a rule can act
A rule can only be switched from flag only to acting when its test set has been run against the rule as it is now, with nothing wrongly flagged. If you change the rule's terms, what it checks or where it runs, run the test set again before setting it to act.
Tick Act without a passing test run to go ahead anyway. That is recorded in the audit log with your name on it.
The trial
Switching a rule to acting starts a trial, 7 days by default — set the length beside the action. During the trial the rule flags as usual and records what it would have done instead of doing it. The rule's card shows how many it would have deleted, timed out, banned or removed, and how many of those flags a moderator dismissed.
Press End trial on the rule when you are happy with it. From then on it really acts. A rule set back to flag only loses its trial.
If a rule runs away
An acting rule that suddenly acts far more in an hour than it usually does pauses itself: more than 10 actions in an hour, and more than 16 times its own hourly average over the last 7 days. Both have to be true. A brand new rule has no average yet, so the first number is the one that stops it, at 11 actions in an hour; a rule that normally acts twice an hour has to reach 33. Neither number can be changed. A paused rule keeps flagging and stops acting, shows Paused on its card with the reason, and appears on the Health page. Press Resume when you have looked at it.
Where a rule runs
Each rule has Channels and Never act on:
- Channels: every channel, only the channels you list, or all but the channels you list. A rule outside its channels does nothing at all there, not even flag.
- Never act on: roles whose members the rule never acts on. Add a Discord role is for a Discord message's author: the rule never deletes their message or times them out. Add a VRChat group role is for the owner of a VRChat profile: the rule never bans or removes somebody who holds that role in your managed group. They are still flagged, and marked Exempt, unless you also tick Do not flag them either, which covers both kinds of role.
A person's group roles are the ones Modbot has on record for them from the last member sweep; a check asks VRChat for nothing. Somebody the sweep has not stored yet has no roles on record and is not spared by a group role, which is the safe way round; members of your team are spared either way. The roles offered in the picker are your group's roles as Modbot last read them, so the list is empty until a group is set and has been read once.
Who set a rule to act
- A rule that acts shows a red badge, such as "Delete, time out 1h" or "Ban from group". Hover over it to see who set it to act. Changing the action records the person who changed it; setting it back to flag only clears the name.
- When several rules match one message, the message is deleted once and the timeout is the longest any of them asked for. When several match one profile, the person is banned or removed once.
- The Discord bot has to be connected, and needs Discord's permissions to delete messages and time out members in your server. A VRChat action needs Modbot's service account to be signed in and able to moderate the group. If an action cannot happen, the flag is still written and the audit log records that it did not happen and why.
Flags, deletions, timeouts, group bans, group removals and dismissals are recorded in the audit log. Changes to rules and to the AutoMod and AI tools cards are recorded there too, as operational entries (seeing them needs See the operational log).
Flags and dismissals
A flag records the rule, the term or topic, what was checked, the person, the Discord message where there is one, the matched words and the reason (a Hub term's note, or the model's sentence).
- No repeats. The same rule and term on the same message, or on the same words in the same part of a profile, is flagged once.
- Dismissing is permanent for that person. Once a flag is dismissed, that rule and term (or topic) never flags or acts on that person again.
- A flag shows the rule as it read when it flagged, even if the rule has been changed or deleted since.
Seeing flags needs See profiles; dismissing them needs Review tickets. See Flags.
Deleting a list or topic keeps the flags it already made.
Sending a flag to Reviews
A flag can go to Reviews, so your team decides it the way it decides everything else.
- Each rule has Open a review for each flag, off to start with.
- Any open flag has an Open a review button, for when you want the team to look at one.
Closing that review asks which way it went. The rule was wrong dismisses the flag; The rule was right confirms it. Both need Review tickets, and both are recorded in the audit log.
The rule's card counts both: confirmations show next to the dismissal rate.
Asking the AI about a flag
With Give an opinion on a flag on, every open flag has an Ask AI button. Modbot sends the model the rule, the words that matched and the text the flag is about (with the earlier messages, for a Discord message), and the model answers Keep or Dismiss with one sentence. The answer shows on the flag and, when the flag has a review, on the review too.
With Propose an action for a flag on as well, the answer also names what a moderator might do. It is advice: nothing is done with it, and the buttons a moderator presses are the same ones as before. Each opinion is one AI request, counted like any other, and needs Review tickets.
Try it
The Try it card checks a piece of text without doing anything. Paste up to 4,000 characters into Text, pick what kind of text it is, and press Try. Tick Include AI topics to run AI topics too.
The answer lists every rule that matched, including rules that are switched off (marked Off), the term and the matched words, and what would happen: Flag only, Delete the message, a timeout, Ban from the group or Remove from the group. Nothing is recorded and nothing is done. Include AI topics appears while AI is on; ticked, the check is a real AI request that uses one of today's AI calls and counts towards spend. If AI topics could not run, the reason is shown under the answer.
Reports
What members tell the mods with /report or the Report to mods menu in Discord, the Reports page where staff read and close them, who sees who reported, and how long the words are kept.
Flags
The Flags page lists what your moderation rules caught in Discord messages and VRChat profiles, filters them by language, and lets you dismiss a flag for good or send it to Reviews.