---
title: "Session channels"
sidebarTitle: "Channels"
description: "The raw HTTP endpoints behind a session's .in and .out streams: append records, read them over SSE, and drain them non-streaming."
---

Every session has two durable streams: `.in` carries records from your clients to the task, `.out` carries records from the task back to your clients. The [`sessions` SDK](/ai-chat/sessions) wraps these as `session.in.*` and `session.out.*`. This page documents the underlying HTTP endpoints for callers that aren't using the TypeScript SDK.

All channel endpoints live under `/realtime/v1/sessions/{session}/{io}`, where:

- `{session}` is the session's friendly ID (`session_…`) or your `externalId`. One token authorizes both forms.
- `{io}` is either `in` or `out`.

Authorize requests with a secret key or a [session public token](/management/authentication#session-scopes). The token's scopes decide what you can do — see [Authorization](#authorization) below.

## Append a record

Append a single record to a channel.

```bash Append to .in
curl -X POST "https://api.trigger.dev/realtime/v1/sessions/{session}/in/append" \
  -H "Authorization: Bearer $TRIGGER_TOKEN" \
  -H "Content-Type: application/json" \
  -H "X-Part-Id: 0f8c2b1e-..." \
  --data '{"type":"user-message","text":"hello"}'
```

The body is the raw record — any text up to 1MiB (records over the per-record cap return `413`). The response is `{ "ok": true }`.

Set the `X-Part-Id` header to a unique value per record to make the append idempotent: replaying the same `X-Part-Id` does not duplicate the record. Appending to a closed or expired session returns `400`.

<Warning>
  Appending to `.out` requires a **secret key**. A session public token (even one with
  `write:sessions`) can only append to `.in` — appending to `.out` with a public token returns
  `403`. The `.out` stream is the task's to write.
</Warning>

## Read a channel over SSE

Subscribe to a channel as a Server-Sent Events stream. New records are delivered as they arrive.

```bash Read .out
curl -N "https://api.trigger.dev/realtime/v1/sessions/{session}/out" \
  -H "Authorization: Bearer $TRIGGER_TOKEN" \
  -H "Last-Event-ID: 42" \
  -H "Timeout-Seconds: 60"
```

| Header | Direction | Description |
| --- | --- | --- |
| `Last-Event-ID` | request | Resume after this sequence number. Set it to the last `id:` you received to pick up exactly where you left off after a disconnect. |
| `Timeout-Seconds` | request | How long the server holds the stream open with no new records before closing, `1`–`600`. |

Each SSE event carries:

- `id:` — the record's sequence number. Use the most recent one as `Last-Event-ID` to resume.
- `data:` — a JSON record `{ "data": <record>, "id": <id> }`. For `.out` on a `chat.agent` session, `data` is a UI message chunk (text, reasoning, tool call, or a custom data part).

```text
id: 42
data: {"data":{"type":"text","text":"echo: hello"},"id":42}
```

### Control records

Some `.out` events are **control records** rather than data. A control record has an empty body and carries a `trigger-control` header naming its subtype:

| Subtype | Meaning |
| --- | --- |
| `turn-complete` | The current turn finished. Carries sibling headers `public-access-token` (a refreshed session token), `session-in-event-id`, and `last-event-id`. |
| `upgrade-required` | The session needs to hand off to a run on a newer deployed version. |

Route control records by their subtype instead of treating them as message content. The TypeScript SDK does this for you — `session.out.read` filters control records out of the chunk stream and surfaces them through `onControl`.

## Drain records non-streaming

Fetch a batch of records without holding an SSE connection open. Useful for polling or for reading a tail at startup.

```bash Drain .out
curl "https://api.trigger.dev/realtime/v1/sessions/{session}/out/records?afterEventId=42" \
  -H "Authorization: Bearer $TRIGGER_TOKEN"
```

Pass `afterEventId` to return only records after that sequence number; omit it to read from the start of the retained window. The response is:

```json
{
  "records": [
    { "data": { "type": "text", "text": "echo: hello" }, "id": 43, "seqNum": 43 }
  ]
}
```

Each record carries `data`, `id`, `seqNum`, and an optional `headers` array (present on control records). Page forward by passing the highest `seqNum` you received as the next `afterEventId`.

## Authorization

The action you can take depends on your token and the channel:

| Action | Endpoint | Required authorization |
| --- | --- | --- |
| Subscribe (SSE) | `GET .../{io}` | `read:sessions:{id}` — works on both `.in` and `.out` |
| Drain records | `GET .../{io}/records` | `read:sessions:{id}` — works on both `.in` and `.out` |
| Append to `.in` | `POST .../in/append` | `write:sessions:{id}` |
| Append to `.out` | `POST .../out/append` | Secret key only |

Reads work in both directions for a `read:sessions` token. Writes split by direction: a `write:sessions` token can append to `.in`, but `.out` is reserved for the task and requires a secret key. See [session scopes](/management/authentication#session-scopes) for how to mint a token.

## Using the SDK instead

If you're writing TypeScript, the [`sessions` SDK](/ai-chat/sessions) is the ergonomic path. `sessions.open(idOrExternalId)` returns a `SessionHandle` whose `session.in` and `session.out` channels call these endpoints for you, with auto-retry, `Last-Event-ID` resume, and control-record routing built in:

```ts Your backend
import { sessions } from "@trigger.dev/sdk";

const session = sessions.open(chatId);

// append to .in
await session.in.send({ type: "user-message", text: "hello" });

// read .out over SSE
const stream = await session.out.read({ signal: AbortSignal.timeout(30_000) });
for await (const chunk of stream) {
  console.log(chunk);
}
```

See [`session.in`](/ai-chat/sessions#session-in-—-clients-→-task) and [`session.out`](/ai-chat/sessions#session-out-—-task-→-clients) for the full handle API.