Modbot Docs
AI

AI settings and providers

Turning AI on, choosing a provider, the address, key and model, the Test button, and what each AI feature sends to the provider.

Modbot's AI features (moderation rules, insights and Chat) all talk to one AI service that you choose. You set it up once, under Settings → AI → Connection, and every feature uses it.

Modbot does not come with an AI service or pay for one. You bring your own: a hosted provider with your own API key, or a model server you run yourself. Everything AI-related is off until you turn it on.

Changing anything on the AI tab needs the Change settings permission (roles and permissions).

Providers

Modbot talks to any service that understands the OpenAI chat format (often called "OpenAI-compatible"). The Connection card offers these presets:

PresetAddress it fills in
OpenRouter (recommended, and the default)https://openrouter.ai/api/v1
xAIhttps://api.x.ai/v1
Anthropichttps://api.anthropic.com/v1/
OpenAIhttps://api.openai.com/v1
CustomNothing. You type the address.

A preset only fills in the Address box. You can edit the address afterwards.

OpenRouter is the one we recommend, and a new Modbot starts on it. One key reaches almost every model there is, which is why Modbot can show you the whole list with prices (see Choosing a model) and why it knows what each request cost. Nothing stops you using another provider, and a Modbot that has already saved one keeps it.

Choose Custom for anything else, including a model server on your own machine or network, such as Ollama (http://localhost:11434/v1). Custom is the only preset that accepts a plain http:// address, and the only one that can be switched on without an API key, because a local server usually needs neither.

Keeping member data in-house

If you do not want profile text, Discord messages or moderation history to leave your own servers, use Custom with a model you run yourself. See What is sent to the provider.

Setting it up

Open the Connection tab

Go to Settings → AI → Connection.

Choose a provider

Click one of the provider buttons. The Address box fills in.

Enter your API key

Paste the key from your provider into API key. Leave it empty for a local server that needs no key.

Choose a model

Press Choose beside Model and pick one from the list, or type a model name yourself. See Choosing a model below.

Test

Press Test. Modbot sends one short message ("Reply with the single word OK.") and shows either the model's name with its answer (for example answered: OK), or the provider's error in its own words.

Turn it on and save

Switch AI on and press Save.

The fields

AI on is the switch for every AI feature at once. With it off, nothing is sent to any provider, whatever the other AI tabs say. With it off you can still save an unfinished form, so you can fill in the address today and the key later. With it on, Modbot will not save until the address, the model and (for any preset except Custom) the key are all there.

Address is the full address of the service, starting with https:// (or http:// for Custom). It cannot contain ? or #, and it cannot contain a user name or password: put the key in the key field.

API key is write-only. Once saved, the page never shows it again; the label changes to API key (stored) so you know one is there. To replace it, type a new one and save. To delete it, press Remove key and confirm.

Changing the address forgets the key

A stored key is only ever sent to the address it was saved with. If you change the address and save without typing a key, the stored key is removed. The same rule applies to Test and the model list: against a different address, they go without the stored key. This stops a key being redirected to another server.

Model is the model every feature uses unless you choose another one for it. Insights and Chat each have their own Model box; left empty, they use this one.

Backup model is optional. When the main model errors, runs out of time, or the provider will not run it, Modbot tries the same request once more on this one. It is picked from the same list as Model.

Modbot does not fall back when a spend limit stopped the call, when the provider refused your API key, or when you closed the page: another model would fail the same way, or spend past a limit you set, or answer a question nobody is waiting for any more. The Test button never falls back either — you are testing one model, and an answer from a different one would not tell you whether that one works.

The call log records both attempts, so you can always see which model actually answered.

Keep the call log for (days) is how long a row in the call log is kept. Thirty days by default. 0 keeps them for ever.

Choosing a model

Every Model box has a Choose button. What it opens depends on your provider.

On OpenRouter

You get OpenRouter's whole list, as Modbot last fetched it, with:

ColumnWhat it is
ModelThe model's name, its id underneath, and chips for what it can do
MakerWho makes it — the part of the id before the /
Input $/1MUS dollars per million tokens you send
Output $/1MUS dollars per million tokens it writes back
Cached $/1MWhat input the provider already has cached costs instead
ContextHow much a request and its answer can hold together, such as "200K"
Per 1,000 callsWhat a thousand calls would cost you — see below

Search by name or id, filter by maker, by Tools, by Structured output, by Free, or by a highest price per million; sort by any column. Click a row to choose that model.

The chips are the short version of what a model can do: Tools, Structured output, Images in, Free. Where a model cannot do something the feature needs, the chip says so instead — No tools on a model Chat could not use, No structured output on one moderation rules could not use — and a router whose price depends on the model it picks shows Price varies.

The order. Models recommended for the box you opened come first, cheapest first. A model is recommended when it can do what that feature needs (tools for Chat, structured output for moderation rules, nothing in particular for insights, which are written as plain text, and both for the Connection box that everything falls back to), when it has a real price, and when OpenRouter added it less than eighteen months ago. Modbot never favours a particular model by name.

Per 1,000 calls appears once Modbot has AI usage of its own to work from: it takes the average request that feature has actually made over the last thirty days — input, cached input and output tokens — and prices a thousand of them on each model, using the same sums as Limits and spend. With no usage yet, the column is not shown.

Fetched in the corner says how old the list is; Refresh fetches it again from OpenRouter, which is the same fetch the Limits tab's Fetch prices button makes. Modbot fetches on its own once a day while AI is on and set to OpenRouter.

On any other provider

Modbot asks your address for its own list of models and shows it, with the price beside any model you entered a price for on the Limits tab, or No price.

A model that is not on the list

You can always type a model id yourself, whatever the list says: the box keeps what you type, and the picker marks it Not in list. Use Test to check it works.

The Test button

Test uses the values on the form, not the saved ones, so you can check a new key before it replaces one that works. It needs an address and a model.

A test is a real request to the provider. It is counted as the Connection test feature on the Limits tab, and a spend limit can stop it. If a limit is reached, Test shows that limit's sentence instead of calling the provider.

If the model list does not load, the address may not offer one. That is not an error: type the model name yourself and use Test to check it.

What is sent to the provider

Everything below goes to the address on Connection, with your key. Text your members wrote — Discord messages, VRChat display names, bios, status lines and pronouns — is sent to the AI provider you choose, and so are pictures when a rule is set to look at them. Your provider's own terms decide what happens to it there.

The first time you switch AI on, Modbot shows you this list and you have to confirm it. That happens once for the deployment, and it is recorded in the audit log with your name, the time and the address you were shown.

FeatureWhat is sent
TestThe single message "Reply with the single word OK."
Model listA request for the model list. No group data.
AutoMod (AI topics and the flag opinion only)Modbot's instructions; the name, "What to catch" text and sensitivity of each AI topic that applies; which kind of text it is (a Discord message, a VRChat display name, bio, status or pronouns); and the text itself, up to 4,000 characters. No separate name or id field is sent, but the text can itself be a name or mention one. Term lists never send anything.
Conversation, when a rule reads itThe messages before the one being checked, in the same channel or thread, and the message it replied to — each with its author's name, clipped to 500 characters. Only for rules set to read the conversation.
Pictures, when a rule looks at themUp to four pictures per check, of six kinds: a Discord message's picture attachments, the author's Discord avatar, and a VRChat profile picture, user icon, profile banner and avatar picture. OpenRouter, OpenAI and xAI are sent a link to each picture; any other endpoint is sent the picture itself, which Modbot downloads first from public addresses only, at 4 MB or smaller. A Discord avatar link contains the person's Discord id. Off unless switched on for that rule.
Opinion on a flag, when a moderator asks for oneThe rule's name and instructions (up to 2,000 characters), the words or picture label that matched and why it matched, the kind of text it was, and the flagged text, up to 4,000 characters. For a rule that read the conversation, the earlier messages go too, each with its author's name. No pictures.
AlertsModbot's instructions and the alert's figures: what was counted, the period it covered, the figure and what is normal, and for the two instance alerts the world's name (its id when Modbot has no name for it). No person's name, id or messages.
InsightsModbot's instructions and the insight's figures: counts for the period and the period before, and short lists of the busiest worlds and instances by world name. No person's name, id or messages.
AI briefs, when a moderator asks for oneModbot's instructions, and the newest audit log entries, up to 100, that an instance's or a person's Activity tab shows the asker: each entry's id, time, what happened, the names of who it was about and who did it (their ids when Modbot has no name), the world and instance, and the entry's own details, which can carry ids and text people wrote. One call, no tools.
ChatModbot's instructions (including the group's name, the current time and any extra instructions you wrote), the whole conversation so far, the list of tools on offer, and every tool result. Tool results are Modbot data: names, ids, bios, history, case files, ban history, audit log entries, Discord messages and members, flags, who is in an instance. The asker's Modbot account id is not sent. After a new conversation's first reply, one more short call sends the first question and the first answer to name it.

The MCP server is different: it is off until you switch it on, it sends nothing to the provider on Connection, and an app a person connects to it receives what the same tools return, as that person.

AI requests go out on their own connection. They never use the proxy you may have set up for VRChat.

Where to next

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