Streaming

Not yet available. The Agent SDK is tracked to be built — today, use the Client SDKs against the live chat completions API. The docs below describe the intended design.

Set stream: true in callModel() to receive token deltas as they arrive. The SDK handles SSE reassembly, tool-call buffering, and result injection — the streaming interface is the same as the non-streaming one, with an added event emitter for incremental output.

Basic streaming#

const result = await client.callModel({
  model: "anthropic/claude-sonnet-5",
  messages: [{ role: "user", content: "Write a haiku about the Sahara." }],
  stream: true,
  onToken: (delta) => process.stdout.write(delta),
});

console.log("\nDone. Tokens used:", result.usage.totalTokens);

The onToken callback fires for each text delta. The returned result is the same TurnResult as non-streaming — output is the fully assembled string, items is complete, usage is final.

Streaming with tool calls#

When the model makes tool calls mid-stream, the SDK pauses the token stream, executes the tool, injects the result, and resumes. You can observe this with onEvent:

const result = await client.callModel({
  model: "google/gemini-2.5-pro",
  messages: initialMessages,
  tools: [weatherTool],
  stream: true,
  onToken: (delta) => process.stdout.write(delta),
  onEvent: (event) => {
    if (event.type === "tool_call_start") {
      console.log(`\n[calling ${event.name}...]`);
    }
    if (event.type === "tool_call_end") {
      console.log(`[${event.name} returned]\n`);
    }
  },
});

Stream events#

Async iteration#

The SDK also exposes an async iterator for environments where callbacks are awkward:

const stream = client.callModelStream({
  model: "meta-llama/llama-4-maverick",
  messages: initialMessages,
  tools: myTools,
});

for await (const event of stream) {
  if (event.type === "token") process.stdout.write(event.delta);
  if (event.type === "turn_end") {
    const result = event.result;
    console.log("\nBalance remaining:", result.balance);
  }
}

Streaming in a multi-turn loop#

stream: true applies per callModel() call. You can mix streaming and non-streaming turns freely — the item list works the same way either way.

Error handling mid-stream#

If the upstream model sends an error event, the SDK emits an error event and resolves callModel() with stopReason: "error". The item list up to that point is still valid and can be used to resume.

Also see#

• Call Model — the full callModel() API.
• Streaming (API) — the underlying SSE protocol.
• Tools — how tools interact with streaming.