A Discord AI chatbot answers your members in natural language. With NoriaFlow it goes further than a chat window: the bot remembers each member, moderates with your own rules, and can act by running your workflows as tools. You build it visually, with your own bot and your own API key.
What you can build
- An assistant behind a
/askcommand that knows your server and answers in the member's language. - An AI moderator that reads each message against your rules and removes scams and abuse, with a reason the member understands.
- A support agent that can open a ticket, give a role or look something up, because those actions are workflows you gave it as tools.
- A Twitch co-host that answers
!aiin chat, in your tone.
The AI Agent step
Every AI template is built on the AI Agent step. It takes a Request (usually the member's message, {{content}} for a message or {{option_question}} for a slash command option) and Instructions that set its role and its limits. It returns its answer as {{text}}, ready to send.
Three options turn a chatbot into an agent:
| Option | What it does |
|---|---|
| Conversation memory | Per member: the agent remembers its recent exchanges with each person. Custom key: one conversation per channel or per ticket. |
| Tools | Up to three of your workflows the agent may run. It fills in the arguments you name, reads the result and answers with it. |
| Answer format | JSON fields: the agent returns named values such as verdict and reason, read as {{data.verdict}}, so the next step can branch on them. |
Build it in five steps
- Save your AI API key. In Connections, add an OpenAI key (origin
https://api.openai.com, Bearer) or an Anthropic key (originhttps://api.anthropic.com,x-api-keyheader). A model on your own machine with LM Studio or Ollama needs no key. - Start from an AI template. The Discord AI Assistant opens with a
/askcommand, the agent and a reply already wired. - Write the instructions. Say who the bot is, how long its answers are, which language it uses and what it must never do. Specific rules beat long ones.
- Give the agent tools (optional). Pick a workflow, name it (
give_role), describe when to use it, and list its arguments (user_id, role_name). The workflow receives them as{{user_id}}and{{role_name}}. - Test, then go live. A test run simulates the AI and every tool, so nothing is sent and no API credit is spent. Publish when the flow looks right.
Moderation that understands context
Keyword filters delete a quote about a scam and miss a scam written with emojis. The Discord AI Moderator template asks the agent for two JSON fields, verdict and reason, with temperature 0 for consistent decisions. A condition checks {{data.verdict}}, and only a clear violation is deleted, with {{data.reason}} explained to the member.
Write your real rules in the instructions, including what is allowed. Keep Discord's own AutoMod on for instant keyword blocks, and let the AI handle what needs judgement.
Safe by design
- The AI only acts through your workflows. It cannot run arbitrary commands; a tool does exactly what you built into it, and every run appears in your history with the tools used and their results.
- No repeated actions. If the AI provider is briefly overloaded, the request is retried without running a tool twice.
- Bounded. The agent has a step limit (5 by default) and must answer on its last step.
- Your key, your data. Keys are encrypted in Connections, and conversation memory expires on the schedule you set.
Beyond Discord
The same agent works on Twitch with the Twitch AI Chat Co-host template, and in your own apps with the AI Answer API template, which turns a workflow into an HTTP endpoint. You can also let ChatGPT or Claude run your workflows directly: see Connect ChatGPT or Claude with MCP.