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

# Vercel AI SDK

> Build a research agent with AgentMuxer and Vercel AI SDK.

In this guide, you'll build an agent that researches the latest TypeScript release and reports what it found, its sources, and the tool cost.

<Accordion title="Add to an existing agent">
  Use your existing `model` and `prompt` with AgentMuxer tools:

  ```typescript theme={null}
  import { agentmuxer } from "@agentmuxer/sdk/ai-sdk";
  import { ToolLoopAgent } from "ai";

  const mux = await agentmuxer();
  try {
    const agent = new ToolLoopAgent({ model, tools: mux.tools });
    const result = await agent.generate({ prompt });
    console.log(result.text);
  } finally {
    await mux.close();
  }
  ```

  The helper returns `{ tools, close }`. It does not choose a model or run the agent.
</Accordion>

## Before you start

* Node.js 22.18 or later with ESM; see [compatibility](/reference/compatibility).
* An [AgentMuxer application key](/sdk/applications) and credits for paid calls.
* An OpenAI API key for the model.

<Note>
  This example asks the agent to spend up to \$0.10 on one tool call. Model charges are separate. Set an [application spending limit](/sdk/applications) if you need a hard cap; the prompt alone doesn't enforce one.

  This example cannot collect payment confirmation. [Configure your account for unattended purchases](/guides/billing#unattended-applications) before running it. This setting also affects your personal agents.
</Note>

<Steps>
  <Step title="Install">
    In a new directory, create an ESM project and install the integration:

    ```bash theme={null}
    mkdir my-agent
    cd my-agent
    npm init -y
    npm pkg set type=module
    npm install @agentmuxer/sdk @ai-sdk/mcp ai @ai-sdk/openai zod
    ```
  </Step>

  <Step title="Set your keys">
    Set both keys in your terminal:

    ```bash theme={null}
    export AGENTMUXER_API_KEY='your-application-key'
    export OPENAI_API_KEY='your-model-provider-key'
    ```

    No browser sign-in is needed. Keep both keys on your server and out of source control.
  </Step>

  <Step title="Create your agent">
    Save this as `agent.ts`:

    ```typescript agent.ts highlight={9,13,18} theme={null}
    import { agentmuxer } from "@agentmuxer/sdk/ai-sdk";
    import { openai } from "@ai-sdk/openai";
    import { ToolLoopAgent } from "ai";

    const prompt = `Use AgentMuxer to research the latest TypeScript release.
    Summarize three changes with sources and report the tool cost.
    Spend at most $0.10 on one tool call. Stop if it costs more.`;

    const mux = await agentmuxer();
    try {
      const agent = new ToolLoopAgent({
        model: openai("gpt-5.5"),
        tools: mux.tools,
      });
      const result = await agent.generate({ prompt });
      console.log(result.text);
    } finally {
      await mux.close();
    }
    ```
  </Step>

  <Step title="Run">
    ```bash theme={null}
    node agent.ts
    ```

    The example asks the agent to print:

    * Three TypeScript release changes.
    * Sources for the summary.
    * The actual tool cost, or the reason it could not complete the purchase.

    To try another task, edit `prompt`. Include the spending you allow.
  </Step>
</Steps>

## Stream the result

In `agent.ts`, replace the `agent.generate()` call and `console.log(result.text)` with:

```typescript theme={null}
const result = await agent.stream({ prompt });
for await (const part of result.textStream) {
  process.stdout.write(part);
}
```

Keep this inside the `try` block. The existing `finally` closes the connection after you consume the stream. For an HTTP response stream, attach cleanup to completion, error, and cancellation; do not close the connection when you merely return the stream.

## Connection options and approvals

The helper accepts `initializationOptions` and `onUncaughtError` and disables automatic MCP tool-call retries. Configure model behavior and tool permissions in your AI SDK agent. See the [SDK reference](/reference/sdk).

In SDK 0.1.0, the helper does not expose a flow for resuming AgentMuxer purchase-confirmation requests. Use unattended purchases as described above or a client with confirmation support. Application spending limits still apply.

<CardGroup cols={2}>
  <Card title="Applications and keys" icon="key" href="/sdk/applications">
    Manage credentials and spending limits.
  </Card>

  <Card title="Run in production" icon="server" href="/sdk/production">
    Handle deployment, failures, and tracing.
  </Card>
</CardGroup>