Modbot Docs
AI

Chat

Ask questions about your group in plain words. Who can use Chat, what it can look up, what is stored, AI briefs of an instance or a person, and its settings.

The Chat page lets a moderator ask questions in plain words, such as "has this person been banned before?", "who is in our instances right now?" or "how many people joined this month?". The AI model answers by looking things up in Modbot with a set of tools, and every lookup is shown beside the answer.

Chat looks things up. It cannot kick, ban, unban, warn or change anything, and there is no tool that could. The one thing it can do is file feedback about Modbot for the admins, and only after an admin switches that tool on and you say yes to its draft (see File feedback). It is also told not to write ban reasons, case files or reports for you, and not to give anyone a score or verdict.

Who can use it

Chat needs three things:

  1. AI set up and switched on under Settings → AI → Connection (AI settings).
  2. Chat on under Settings → AI → Chat. It starts off.
  3. The Use AI chat permission on the person's role (roles and permissions).

Use AI chat is not part of the built-in Moderator or Viewer roles. Using Chat sends group data to your AI provider, so give it on purpose. Administrators have it.

The Chat link appears in the sidebar, the phone menu, the command palette and on g c only when all three are in place: AI on, Chat on, and Use AI chat on the person's role. If Chat or AI is switched off, the link is gone, but a direct link to /chat still opens the page, which says Chat is off. Someone who already has conversations still sees them and can read them, but the message box is greyed out.

Answers only use what the asker can see

Every tool runs with the permissions of the person asking. A tool is only offered to the model if the asker holds the permission it needs, and a tool that was not offered never runs, even if the model asks for it. Where a page in Modbot shows different things to different people (the audit log, who is in an instance), the Chat tool uses the same rules as that page, with the asker's permissions.

So two moderators asking the same question can get different answers, and Chat never reveals something the asker could not already open in Modbot.

Tools

ToolWhat it looks upNeeds
Find a personVRChat users Modbot has a profile for, by id or part of a display name (up to 10)See profiles
Open a profileOne user's stored profile: display name, bio, status, pronouns, when they joined VRChat, the 18+ mark, when Modbot first and last saw them, and group membership and ban standingSee profiles
Person's historyThe most recent audit log entries about one person and any Discord or Modbot account tied to them, everything or only the moderation history, presence or DiscordSee the audit log
Person's case filesCase files (ban write-ups) about one personSee profiles
Person's bansWhether one person is banned now or was before, with every ban and unban and who did itSee the audit log
Search membersThe group's member list, by name or id, with roles, join date and last seenSee members
Search the audit logAudit log entries, filtered by type, person, who did it and datesSee the audit log
Live instancesThe group's open instances right now, with head counts, which moderators are watching and who is thereSee live instances
Find a worldWorlds Modbot has seen, by id or part of a name (up to 10)See analytics
Open a worldOne world: name, author, capacity, time people spent in it and its recent instancesSee analytics
Open an instanceOne instance: where and when it ran, how busy it was, who was in it and what was recorded thereSee analytics
Group figuresThe group's figures over a number of days (30 unless asked otherwise, up to 365): member count, joins, leaves, invites, join requests, roles and how long members have been membersSee analytics
Server figuresThe Discord server's figures over a number of days: members, joins, leaves, messages, minutes in voice, active members, busiest channels, who wrote most, bans, kicks, timeouts and messages removedSee analytics
Search Discord messagesStored Discord messages, by who wrote them, the channel, words in them and dates. Deleted ones come back marked deletedRead Discord messages
Around a messageThe messages sent just before and just after one message, in the same channel or threadRead Discord messages
Open a Discord memberOne member of the Discord server: names, roles, when they joined, whether they left or are timed out, and their linked VRChat accountSee members
Search Discord membersThe Discord member list by name or id, and everyone holding one roleSee members
Who left DiscordThe people who most recently left the Discord serverSee members
Account linkWhether someone's VRChat and Discord accounts are linked, the other account, and links that were endedSee profiles
Person's flagsOne person's AI moderation flags: which rule matched, what it matched, what was done, and whether a moderator dismissed itSee profiles
Recent flagsThe most recent AI moderation flags across the group, and how many are still openSee profiles
What's onCalendar events over a window of days: when each runs, its world, whether it repeats and its stateSee calendar
Open an eventOne calendar event: its description, world and instance settings, where it was published and whether Modbot opened an instance for itSee calendar

A list of results is capped, and the answer is told when there were more. Searching Discord messages on words alone looks at the last 90 days unless dates are given.

All of these only read. Anyone with Change settings can switch any of them off for everyone in the Chat settings.

File feedback

ToolWhat it doesNeeds
File feedbackFiles a bug, an idea or a change as feedback for the admins, as the person askingSend feedback

This is the one tool that acts, so it starts off. Switch it on under Settings → AI → Chat → Tools. Once it is on, Chat is told to offer it: when you describe a bug, something Modbot is missing or a change you wish for, it offers to file it, shows you a draft (the kind, a title and the details, with names of people, worlds and groups taken out) and files it only after you say yes. The item is saved on your own Modbot server and shows on the Feedback page with your name as the sender. Nothing is sent to Modbot's makers or to GitHub.

Using the page

Asking

  • Type into the box at the bottom and press Enter to send; Shift+Enter starts a new line. A message can be up to 4,000 characters. The box grows as you type. AI settings names the model in use.
  • In Chrome and Edge a microphone button sits beside Send. Click it and speak; what you say fills the box after any text already there, and the button pulses while it listens. Click it again to stop. Nothing is sent until you press Send. It listens in your browser's language. If the browser's microphone is blocked, the box says Microphone blocked. Firefox has no speech recognition, so there is no button there. The browser sends the sound to its maker's speech service; Modbot never receives it.
  • Pressing / anywhere on the page puts the cursor in the message box.
  • The answer is written a few words at a time. The page follows it down unless you scroll up to read something, and then a Jump to latest button appears.
  • While the model is working, Send becomes Stop. Stopping ends the reply at the provider as well as on screen; the part that was already written is kept and marked Stopped.

Reading an answer

  • Answers are written in Markdown: headings, lists, tables and code, with a copy button on each code block. Nothing the model writes can become a link out of Modbot.
  • Every lookup is a step inside the answer, with a tick when it worked. Open one to see what it was asked and what came back in plain words, or press Raw for the result as the model got it. Several lookups in a row are grouped as N steps.
  • People, worlds and instances the lookups found are underlined in the answer and shown as buttons under the step. Either opens the usual popup.
  • Sources under the answer are the rows it was built from. Each one opens where it lives: a person, world or instance in its popup, an audit log entry in the audit log at that entry, a case file on its own page, a Discord message in that person's Messages tab with the message marked, and an event on the calendar. The model is told to answer only from what the lookups returned and never to write an id it was not given.
  • Hovering over an answer shows Copy and Try again; hovering over your last question shows Copy and Edit.

Asking again, and editing

  • Try again writes another reply to the same question. Edit sends a changed question from that point.
  • Nothing is thrown away: both versions are kept, and ‹ 1/2 › beside the message moves between them. Moving back to an older version shows the reply that followed it.

Conversations

  • The list on the left is grouped into Today, Yesterday, Previous 7 days and Older, and can be searched. New chat starts a fresh one, and the list can be hidden. On a narrow screen it opens as a panel over the page.
  • After the first answer, the model writes a short name for the conversation. The ⋯ menu on a row renames it or deletes it; deleting asks first.
  • Each conversation has its own address, such as /chat/0199…, so it can be bookmarked or reopened. It is still only yours to open.
  • The coin button in the header shows what the whole conversation has used and cost, every version of it included.
  • A conversation holds up to 200 messages (tool calls, their results and older versions count). When it is full, start a new one.

A reply can end early:

  • The reply used every tool call it was allowed. The model hit the tool call limit and answered with what it had.
  • The reply took too long. It passed the time limit.
  • The AI provider could not answer. The provider refused or could not be reached.
  • A spend limit's sentence, such as Your daily AI spend limit is reached. Limits are checked before a reply starts, never in the middle of one. See Limits and spend.

Conversations are private

A conversation belongs to the account that started it. Nobody else can list, open, continue or delete it, administrators included; to anyone else it looks as though it does not exist. Conversations stay until their owner deletes them. Staff accounts are never deleted, only disabled, so a disabled account's conversations stay where they are and nobody else can read them. See Staff accounts.

Conversations are not part of the audit log. Modbot's server log records who used Chat, which tools ran, whether they worked and how long they took, but never the text of a message or what a tool was asked.

Asking about a person is recorded

A question whose lookups read about people records one audit log entry naming you and everyone the lookups returned: Wren asked about TeaSpoon in chat. It appears in the audit log and in each of those people's own history, exactly like any other lookup, so asking through Chat is as visible as opening a profile.

One entry per question, however many lookups it made. Try again and Edit send a question, so each records again. Your question is never stored in the entry — only who asked, who was read, which lookups read them and which conversation it was. An AI brief records the same entry, reading for an AI brief.

What leaves Modbot

Each reply sends your AI provider Modbot's instructions (with your group's name, the current time and any extra instructions), the conversation so far, the list of tools on offer, and every tool result. Naming a new conversation is one more short call, which sends the first question and the first answer; it is counted and limited like every other call Chat makes. If it fails, the conversation keeps the first words of the question as its name. Tool results are real group data: names, bios, ban history, audit log entries, who is in a instance. The asker's Modbot account id is not sent.

If that data must not leave your own servers, use a model you run yourself. See AI settings.

AI briefs

An AI brief is a short summary of what the audit log recorded about one instance or one person, for a moderator arriving late or looking back after an event. It is in the popups:

  • Summarise what Modbot recorded, at the top of an instance's Activity tab.
  • Brief me, at the top of a person's Activity tab.

A brief states what happened and when, oldest first, one line per thing. It gives no opinions, no recommendations, no guesses about why anybody did anything, no scores or verdicts, and never writes ban reasons, case files or reports. The model is told all of that. Each line ends with the audit log entries it rests on, such as [#1234, #1240]; an id that was among the entries sent opens the audit log at that entry, and an id the model made up stays plain text. The line under the title says what the brief was written from: Written by AI from 42 audit log entries, 1 Oct 2026 14:02 to 1 Oct 2026 18:40 (Europe/London). Times are in your own time zone.

What it is built from

The same entries the Activity tab shows you, read with your own permissions, so a brief never says anything you could not open yourself:

BriefEntries sent
InstanceThe newest 100 audit log entries recorded in that instance between the time it opened and the time it closed
PersonThe newest 100 audit log entries about or by any of their VRChat, Discord and Modbot accounts

With nothing recorded, the brief says Nothing recorded. and no call is made.

What leaves Modbot for a brief

One call to the Chat model, with no tools: Modbot's instructions, and for each entry its id, time, what happened, the names of who it was about and who did it (their ids where Modbot has no name), the world and instance, and the entry's own details, which can carry ids and text people wrote, such as a note or a reason. For an instance, the instance's name and when it ran; for a person, their name.

What it counts as

A brief is a Chat call. It uses Chat's model and time limit, and it is counted and limited as Chat: Chat's spend limits, the limits on the person and their roles, and the person's monthly allowance all stop it, with the same sentence Chat gives. It is a row in the call log under Chat, with the whole prompt and answer kept, because somebody pressed a button for it. A brief that read about people records Wren asked about TeaSpoon for an AI brief in the audit log, as a Chat question does.

Keeping one

Copy copies the brief with its last line. On a person's brief, Save as note saves it, once, in their notes, marked AI brief: exactly the brief's words and its last line, written by Modbot.

Who can use it

Briefs need AI on, Chat on and AI briefs on under Settings → AI → Chat (off when Modbot is installed), and Use AI chat on the person's role. The instance brief also needs what the instance's Activity tab needs: See analytics and See the audit log.

Chat settings

Settings → AI → Chat needs Change settings. It has three cards; Save on any of them saves them all.

Chat

  • Chat on switches the Chat page on for everyone with Use AI chat. The card shows AI is off on Connection. when AI itself is off.
  • AI briefs on offers AI briefs in the popups to everyone with Use AI chat, while Chat is on too. It starts off.
  • Model is the model Chat uses. Left empty, Chat uses the model on Connection (shown as the placeholder).
  • Extra instructions are added to the end of Chat's instructions, under a heading saying they come from your group's operator. Use them for things like your group's own words for roles, or how short answers should be. Up to 8,000 characters. They are not sent with an AI brief, which has fixed instructions of its own.

Limits

SettingDefaultAllowedWhat it does
Steps per reply80 to 50How many lookups one reply may make. At 0, no tools are offered at all.
Reply length (tokens)2,000256 to 32,000The most the model may write in each step of a reply, in tokens (small pieces of text, see Limits and spend).
Time limit (seconds)12010 to 600How long one whole reply may take, lookups included.

These limit a single reply. To limit what Chat costs, use spend limits, which can be set for everyone, for Chat, for a role or for one person.

Tools

One switch per tool. Beside each is Reads or Acts and the permission it needs. Read-only tools start switched on; File feedback, which acts, starts off. Switching a tool off removes it for everyone, whatever their permissions.

Modbot is not endorsed by or affiliated with VRChat Inc. or Discord Inc. VRChat is a trademark of VRChat Inc. Discord is a trademark of Discord Inc. SteamVR is a trademark of Valve Corporation. Windows is a trademark of Microsoft Corporation. Docker is a trademark of Docker, Inc. PostgreSQL is a trademark of the PostgreSQL Community Association of Canada. Railway is a trademark of Railway Corp. Other names are trademarks of their respective owners; see Legal. Modbot and its logo are trademarks of the Modbot project. Open source under AGPL-3.0.

On this page