> ## Documentation Index
> Fetch the complete documentation index at: https://docs.syvon.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Chat Stream

> Consume the agent chat endpoint. NDJSON events, MDX replies, threading, per-visitor rate limiting, and billing behavior.

Agent chat is a POST that returns an **NDJSON stream**: one JSON event per line. The SDK yields each parsed event as a `ChatStreamEvent`.

## Assemble the reply

```ts theme={null}
let reply = '';
for await (const ev of client.brain.chat(agentId, {
  prompt: 'What do you make?',
})) {
  if (ev.type === 'error') throw new Error(ev.message);
  if (ev.text) reply += ev.text;
}
```

The assembled reply is **MDX**: paragraphs, headings, and occasional embedded components. Render it with an MDX-capable renderer; do not treat it as plain text.

## Threading

Pass `threadKey` to resume a thread and `history` to supply prior turns explicitly:

```ts theme={null}
const events = client.brain.chat(agentId, {
  prompt: 'More detail on the second one',
  threadKey: threadKeyFromEarlierTurn,
  history: [
    { role: 'user', content: 'List your services' },
    { role: 'assistant', content: earlierReply },
  ],
});
```

## Scoping

`section` scopes the turn to a page or section of the agent's site (useful for contextual chat on a specific page). `brandSlug` selects among a multi-brand agent's brands.

## Rate limiting and billing

* The agent's **owner is billed** for every completed turn. Your integration is spending someone else's credits; keep that in mind before auto-firing prompts.
* Pass `client`, a **salted visitor hash your app mints** (never a raw IP or email), to get per-visitor caps.
* Without `client`, the whole workspace shares a looser backstop.
* A rate-limited turn returns `429`. Back off and retry; do not retry-loop against it.