AI Agent

Comes with strAImer: this module is installed together with the app, so there is nothing to download. It is listed under Module Catalog → Built-in.

What this page is

Chat Bot → AI Agent is where the bot's behaviour is set: when it speaks, what it sounds like, what it is allowed to look up, and how much it remembers. What it knows about you lives on a separate page, Chat Bot → Knowledge Base.

Every answer costs AI credits. The replies are generated on the strAImer server, so the agent needs an account, and each reply is charged. The account is free and comes with 100 credits every month at no cost — the page says so itself if you are not signed in yet, under Sign in to use AI, and its Sign in button goes to Settings → strAImer Account. Settings on this page are saved whether you are signed in or not; they simply start working once you are.

How to set it up

1. Pick a model

General Settings → AI Model. Each model in the list carries its price as {n} cr. / answer, so the choice is a price you can read before making it. A model your plan does not cover is marked requires the {tier} plan; while you are signed out the list says Sign in to choose a model, and a model saved earlier that is no longer offered is marked saved earlier.

There is no API token field here. The account and its key live in Settings → strAImer Account — the note on the page says as much: "This page is only about bot behavior."

2. Give the bot a name

Trigger Settings → Bot Nicknames in Chat, comma-separated. The agent answers only when it is addressed, and this field is empty on a fresh install — there are no built-in names, not "bot", not "ai", not "straimer". Until something is in it, the agent is silent no matter what else is set up.

A connected bot account counts as a name automatically, and so does a direct @mention — but the bot account only counts if one is actually set up under Connections → Bots.

The name has to open the line. "bot, what do you think" reaches the agent; "hey, what does the bot think" does not.

3. Decide how often it speaks

Three settings in the same block, doing different jobs:

• Only while streaming — on by default. While you are not live the bot answers nobody in chat, so viewers cannot spend your credits between streams. Your own messages from the channel account always get an answer, so you can test it from the Twitch or Kick chat. Donations and voice commands are not affected. It relies on the Stream Status module to tell whether you are live: if that module is off, or if no platform and no OBS reports the state of the stream, the bot answers as before. After a stream ends it keeps answering for about three minutes, so a short connection drop does not count as the end of the stream.
• Chat Response Rate (%) — the share of addressed messages it answers at all. Below 100 some mentions are ignored on purpose.
• Cooldown Between Responses (sec) — ten by default. In a busy chat this is what keeps the bill down: mentions arriving inside the cooldown are skipped, not queued.

Both make the bot look like it is ignoring people, and both are doing exactly what they were set to do.

4. Set its character

System Prompt (Instruction) is the agent's character, and it applies to chat, donations and voice alike unless one of those overrides it. The default is short and safe: "You are a friendly stream assistant. Keep replies short, helpful, and supportive of the streamer."

Write it as instructions to a person about to speak on your behalf. It is the one setting that changes how every reply sounds.

5. Donation replies, if you want them

Donation Responses, and it is off by default because each donation would be a paid request. Switching it on is not enough on its own — three things decide whether a reply actually happens:

• Send replies to — donations do not belong to any chat, so you choose which chats receive the thank-you. Nothing chosen, nothing sent.
• React to donations from — a checkbox per donation service. An unticked source gets no reply; a service installed later is answered automatically.
• Minimum donation amount — counted in your base currency, so 100 in one currency and 100 in another are not treated alike. Zero replies to every donation.

Donation Prompt is optional and inherits the system prompt when empty. It is worth filling in for one reason: a donation reply addresses the donor without an @, because the donation came from a donation service and that person may not be in your chat at all.

6. Memory

Memory Settings (SQLite) decides how much of the conversation the agent carries. Enable Dialog Memory turns it on; Chat Context Window Size is how much of the room's recent chat it sees, and Personal Dialog History Size how much of its history with one particular viewer. Both are capped at 20 messages, and the page prints that cap. How many of them reach the model depends on your plan: on smaller plans the oldest are dropped first.

Clear chat memory after stream wipes the shared chat context about three minutes after the broadcast ends — the next stream does not need yesterday's room. Personal dialogs survive it. Clear memory now wipes it on demand and asks twice: the button turns into Sure? Click again.

7. Web search, and the brake on it

The agent does not know what is happening in the world unless it looks it up. Asked a date-bound question — news, weather, prices, a match result, what holiday it is — it says it cannot check that right now rather than naming something. An invented answer in chat is worse than no answer.

Allow web searches changes that, and a search costs 10 credits against the 1–2 of an ordinary reply. It works differently on the two sides:

• In chat, a message searches only if it contains one of the Search trigger words — "search, google, look up, find online" by default, and yours to change.
• By voice, the agent decides for itself.

Smart mode hands the chat decision to the agent too, instead of waiting for trigger words. The page states the price plainly: "each one costs 2 credits instead of 1, and a web search still costs 10."

Because the agent is now the one deciding, there is a brake beside it. Limit web searches with Searches per hour counts over the last rolling hour, not from the top of the hour. Zero means no searches at all — to remove the limit, switch the toggle off rather than setting zero. Left off, nothing limits the searching.

8. Advanced Prompt Templates

A collapsed block at the foot of the page holding the service text sent to the model: language rule, reply format for chat and for voice, the event lines, and how the knowledge base is introduced. Do not remove the placeholders in curly braces — they are where the real values land.

A template you edit gets a modified badge and stops following app updates. That is the part worth knowing: an improvement shipped in a later version will not reach a template you have customised. Reset to defaults restores every template in the block at once, updates included, after Save Settings; to restore only one, empty its field and save.

What the agent knows without searching

Everything about you and your stream comes from the app itself, costs no search and adds no credits:

• what you are playing now, the stream title, and how long you have been live;
• your profile, links and gear from Chat Bot → Knowledge Base;
• the chat commands that are switched on and open to Everyone, up to the first 40 on your list, so "what commands do you have" is answered from the app rather than guessed. A command limited to Subscriber, VIP, Moderator or Broadcaster is not given to it, and when you have more than 40 the log file records how many were left out;
• what was said earlier in the conversation, as far back as the memory windows reach.

Links deserve a line of their own: the agent sees your whole list and shares only the one that was asked for, in the exact form you wrote it. It will not dump the list, and it asks which one you meant rather than guessing.

A viewer's personal dialogue is kept while they keep talking and is erased after thirty days of silence. Nothing to set and nothing to press: the sweep runs by itself, and only the personal thread goes — what the agent knows from your knowledge base stays.

Which language it answers in

The language of a reply is chosen by the app, not by the character. A chat message is answered in the language it is written in, out of the six languages of the app: Cyrillic text gets Russian, Latin text gets English, Spanish, Portuguese, German or French, judged by common words, and English when the words do not tell. A request by voice is answered in the language the recognition heard: the one set in Language you speak on Voice → Speech recognition, or the one it detected under Detect automatically. A message with no letters at all, such as a donation without text or a line of emotes, gets the language of the app's interface.

That rule goes to the model after the System Prompt and outranks it, so "always answer in English" written in the System Prompt does not hold. To fix the language for good, open Advanced Prompt Templates, change Language rule to something like "Reply in English." and press Save Settings. Like any edited template it is then marked modified and no longer follows app updates. To bring back the automatic choice, empty the Language rule field and save; Reset to defaults would also reset every other template in the block.

What it cannot do

• It cannot work signed out. Answers are generated on the strAImer server; without an account nothing is generated, however the page is filled in.
• It cannot answer for free. Every reply is charged, and a search is charged more. Running out is not silent — the bot says so: "Out of AI credits. Top up in your strAImer account and I am back."
• It has no character picker. Character is the System Prompt, and nothing else on the page changes how it sounds.
• It cannot speak by itself. Whether its chat answers are read aloud is the AI agent answers switch on Voice → Voice Over, under What the bot writes in chat, and they are read with the Bot in chat voice on Voice → Voices. Its answers to your voice are spoken only with Read answers out loud on Chat Bot → Voice control, with the Bot replies voice. This page only produces the text.
• It does not take voice commands here. What the bot may do when you speak to it is on its own page, Chat Bot → Voice control, and the microphone that hears you is set up in Voice → Speech recognition.
• It takes no actions from chat. A clip, a poll, a prediction, a scene, the title or a control panel button are done only when you ask by voice. In chat the bot only talks, to viewers, moderators and your own account alike.
• It cannot start a conversation. Nothing here makes the bot speak unprompted — it answers when addressed, and replies to donations if you switched that on.
• Replies are cut to fit chat: 190 characters on YouTube, 400 elsewhere. A long answer is shortened rather than split across messages. When one reply goes to several chats at once, the shortest limit among them applies to all of them — a donation thank-you sent to YouTube and Twitch together is 190 characters in both.
• Its knowledge is not edited here. Profile, aliases, links, gear and facts are all on Chat Bot → Knowledge Base.