Quickstart
Create an agent, give it something to read, and put it on your site.
Five steps, start to finish. Everything here happens in the dashboard — no code until the
last step, which is one <script> tag.
1. Create an agent
From Agents, click New agent. Pick a template if one fits what you are building, or start blank. A new agent begins as an unpublished draft.
Give it a name your team will recognise. The name is internal; what visitors see is the display name on the Persona tab.
2. Give it a persona
Open the agent and work through the Persona tab:
- Display name and avatar — what appears at the top of the chat.
- System prompt — the instruction the model reads before every conversation. Say what the agent is for, what it should not do, and what to do when it doesn't know.
- Welcome message and suggested prompts — the first thing a visitor sees.
- Fallback message — what it says when it has nothing.
On the Model tab, pick a provider and model. Groq is the default and is free to start with. Leave temperature low (around 0.2–0.4) for a support agent: high temperature makes an agent that improvises, which is the opposite of what you want near your customers.
3. Add knowledge
Go to Knowledge and create a knowledge base, then add documents — upload PDFs, Word files, spreadsheets, Markdown or text, or paste a URL.
Each document is queued, extracted, split into chunks and embedded. Watch the status column until it reads ready. A document stuck on processing usually means the embedding provider is unreachable; failed will tell you why.
Then, back on the agent's Knowledge tab, switch RAG on and select the knowledge base.
4. Test it in the playground
The agent's Playground runs the draft, not the published version, so you can break things safely. Ask it the questions your customers actually ask. Check that:
- answers cite a document, rather than sounding plausible and citing nothing,
- it declines gracefully when you ask something outside its knowledge,
- the tone matches what you wrote in the system prompt.
When it holds up, click Publish. Publishing promotes the draft to the live version.
5. Put it on your site
On the Channels tab, set the widget's colour, position and launcher text, then copy the
embed snippet. Paste it just before </body>:
<script
src="https://YOUR_WEB_HOST/widget.js"
data-agent="pk_your_agent_public_key"
data-api="https://YOUR_API_HOST"
defer
></script>data-agent is the agent's public key. It is safe in client-side HTML — it grants access
to that one published agent's public chat and nothing else.
Reload your site. A launcher bubble appears in the corner.
What to do next
- Install the widget — theming, the JavaScript SDK, CORS.
- Connect a channel — Telegram, WhatsApp, Slack and the rest.
- Human handoff — so a person can step in.
- Automations with n8n — let the agent do things, not just say them.