Getting started
Quickstart
Create your first agent, give it knowledge, and publish it — in a few minutes.
1. Create an agent
From your dashboard, click + New agent and give it a name. That's the only thing required up front — everything else (model, prompt, knowledge, tools, channels) is configured afterward from the agent's own tabs.
2. Add some knowledge
Open the agent's Knowledge tab and add a source:
- Upload — drag in a PDF, Word doc, spreadsheet, HTML, text/Markdown, CSV, JSON file, or a scanned image (OCR'd automatically). Up to 20 MB per file, multiple files at once.
- Paste text — give it a title and paste content directly, useful for a quick FAQ or policy you don't have as a file yet.
New sources ingest in the background — the table shows status live and updates every couple of seconds while anything is processing. See Adding knowledge for the full walkthrough.
3. Check the system prompt and model
Open Settings. The default model (Claude Haiku 4.5) is a good starting point for most agents. The system prompt should describe tone and behavior ("You are a friendly support agent for Acme...") — facts and policies belong in Knowledge, not here. There's a Rewrite with AI button if you want a rough draft expanded into a fuller prompt. See Choosing a model and Writing a system prompt.
4. Try it in the Playground
Every agent page has a ▶ Test button in the header. Open it, ask the kind of question a real customer would ask, and check the Activity trace panel alongside the reply — it shows exactly which knowledge chunks were retrieved and which tools ran. This is the fastest way to catch a gap in your knowledge base before a real customer does. See Playground & execution trace.
5. Publish to a channel
Open the Channels tab. The Website widget channel needs no setup —
it's created instantly and gives you a copy-paste <script> snippet to drop
into your site. Telegram, WhatsApp, Instagram, and Email each need a
connected account first (see Channels).
6. Watch it work
Use Analytics (per-agent or organization-wide) to see conversation volume, grounded-answer rate, and response latency, and use Ops to step in and take over a conversation by hand whenever you need to.
That's the whole loop: knowledge in, model configured, published to a channel, and visibility into what's happening. Everything past this point in the docs goes deeper on each piece.