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

# Streaming Inference

> Stream model responses with SSE.

For real-time output, stream a model response with `sdk.inference.responses.stream()`. You receive typed events as the model generates — text fragments, tool calls, and a final done event. This is raw model output over SSE, with no agent runtime in between.

## Stream a response

```typescript theme={null}
const stream = sdk.inference.responses.stream({
  model: "gpt-4o",
  input: "Write a haiku about state channels.",
});

for await (const event of stream) {
  switch (event.type) {
    case "text-delta":
      process.stdout.write(event.delta);
      break;
    case "tool-call":
      console.log(`\n[Tool call: ${event.name}]`);
      break;
    case "done":
      console.log(`\nTokens: ${event.usage.total_tokens}`);
      break;
  }
}
```

## Event types

| Event        | When                        | Key fields             |
| ------------ | --------------------------- | ---------------------- |
| `text-delta` | Model emits a text fragment | `delta`                |
| `tool-call`  | Model requests a tool call  | `name`, `arguments`    |
| `done`       | Stream completed            | `usage` (token counts) |
| `error`      | Stream failed               | `error`                |

Concatenate `text-delta` fragments to build the full response. The `done` event carries the final usage and receipt.

## Inference streaming vs agent streaming

Both use SSE, but they serve different layers:

|             | Inference streaming               | Agent streaming                                                                     |
| ----------- | --------------------------------- | ----------------------------------------------------------------------------------- |
| Source      | Raw model output                  | Agent runtime (model + tools + memory)                                              |
| Events      | `text-delta`, `tool-call`, `done` | `text-delta`, `tool_start`, `tool_end`, `plan.*`, `swarm_child_*`, `done`, and more |
| Tool events | Tool call requests from the model | Full tool execution lifecycle (start + end + result)                                |
| Plan / goal | Not available                     | `task.*`, `plan.proposed`, `action` events                                          |
| Payment     | Settled on the receipt            | Settled on the receipt                                                              |

Use inference streaming when you want direct model output with no orchestration. Use agent streaming when you want tools, memory, plans, or multi-agent delegation.

## Related

* [Inference](/sdk/inference/overview) — all inference capabilities
* [Agent Streaming](/sdk/streaming/overview) — the richer agent event stream
* [Stream Protocol](/sdk/streaming/protocol) — every agent event type