OpenRouter

Provider through Vercel AI SDK

Install @openrouter/ai-sdk-provider@3.1.0 on the server with the current AI SDK. This route changes the model provider, not Ghost's runtime. Use your configured model ID and a server-only key.

123456import { createOpenRouter } from "@openrouter/ai-sdk-provider";
import { streamText } from "ai";
export function streamOpenRouter(apiKey: string, modelId: string, prompt: string, signal: AbortSignal) {
  const provider = createOpenRouter({ apiKey });
  return streamText({ model: provider.chat(modelId), prompt, abortSignal: signal }).toUIMessageStreamResponse();
}

Return that response from your authenticated route. Use the Vercel React recipe unchanged. Validate input and own history/tool policy in your server. See OpenRouter's provider docs.

Agent SDK with native items

OpenRouter also provides an actual Agent SDK, @openrouter/agent@0.11.0. It is a separate integration path. getItemsStream() emits replacement items, not text deltas; replace the item with the same ID without moving its first position.

12345678910111213141516import type { OpenRouter } from "@openrouter/agent";
import { normalizeOpenRouterItems } from "@ghost-ui/core/chat-normalizers";
function itemId(item: unknown): string {
  if (typeof item !== "object" || item === null) throw new TypeError("Invalid item.");
  if ("id" in item && typeof item.id === "string" && item.id) return item.id;
  if ("callId" in item && typeof item.callId === "string") return `result:${item.callId}`;
  throw new TypeError("A durable item ID is required.");
}
export async function* openRouterMessages(client: OpenRouter, model: string, input: string, signal: AbortSignal) {
  const result = client.callModel({ model, input, signal });
  const items = new Map<string, unknown>();
  for await (const item of result.getItemsStream()) {
    items.set(itemId(item), item);
    yield normalizeOpenRouterItems([...items.values()], { getId: itemId });
  }
}

Run this on the server or an existing authenticated application owner; deliver its canonical snapshots to the plain-data React view or an external store. Keep the native conversation state for follow-up runs; the projected UI messages do not replace it.

Tools, usage and cancellation

A completed function_call item means its input is ready, not that execution succeeded. A function_call_output is a separate result. Tool execution and approval policy stay with the Agent SDK/application configuration. Unknown items remain explicit. Reasoning summaries require deliberate classification.

The items stream excludes usage metadata. Use native usage/full-response APIs when your application needs cost or round accounting; do not invent totals from displayed text. Pass cancellation through the native signal and preserve its outcome. See Agent item streaming and stability.