How to integrate Openai MCP with Vercel AI SDK v6

Trusted by
AWS
Glean
Zoom
Airtable

30 min · no commitment · see it on your stack

Openai logo
Vercel AI SDK logo
divider

Introduction

This guide walks you through connecting Openai to Vercel AI SDK v6 using the Composio tool router. By the end, you'll have a working Openai agent that can list all available openai models, upload a file for fine-tuning, create a new assistant with gpt-4 through natural language commands.

This guide will help you understand how to give your Vercel AI SDK agent real control over a Openai account through Composio's Openai MCP server.

Before we dive in, let's take a quick look at the key ideas and tools involved.

Also integrate Openai with

TL;DR

Here's what you'll learn:
  • How to set up and configure a Vercel AI SDK agent with Openai integration
  • Using Composio's Tool Router to dynamically load and access Openai tools
  • Creating an MCP client connection using HTTP transport
  • Building an interactive CLI chat interface with conversation history management
  • Handling tool calls and results within the Vercel AI SDK framework

What is Vercel AI SDK?

The Vercel AI SDK is a TypeScript library for building AI-powered applications. It provides tools for creating agents that can use external services and maintain conversation state.

Key features include:

  • streamText: Core function for streaming responses with real-time tool support
  • MCP Client: Built-in support for Model Context Protocol via @ai-sdk/mcp
  • Step Counting: Control multi-step tool execution with stopWhen: stepCountIs()
  • OpenAI Provider: Native integration with OpenAI models

What is the Openai MCP server, and what's possible with it?

The Openai MCP server is an implementation of the Model Context Protocol that connects your AI agent and assistants like Claude, Cursor, etc directly to your OpenAI account. It provides structured and secure access to your models, assistants, files, threads, and fine-tuning jobs, so your agent can perform actions like managing assistants, handling conversations, uploading or organizing files, and working with OpenAI models on your behalf.

  • Assistant and conversation management: Quickly create, update, or delete OpenAI assistants and manage threads or messages for seamless conversational flows.
  • File uploads and organization: Let your agent upload new files, list all uploaded documents, or delete unnecessary files to keep your workspace tidy.
  • Model discovery and utilization: Effortlessly list all available OpenAI models—including vision and multimodal—and retrieve their details to choose the best fit for your tasks.
  • Fine-tuning job insights: View a complete list of your organization's fine-tune jobs and track their progress or review results as needed.
  • Thread and run management: Create, modify, or inspect threads and run steps to fully control and monitor interactive agent conversations.

Supported Tools & Triggers

Tools
Add Upload PartTool to add a part (chunk of bytes) to an Upload object.
Cancel batchTool to cancel an in-progress batch.
Cancel evaluation runTool to cancel an ongoing evaluation run.
Cancel ResponseTool to cancel a background model response by its ID.
Cancel RunTool to cancel a run that is currently in progress.
Cancel uploadTool to cancel an upload.
Compact ResponseTool to compact a conversation or response to reduce token usage.
Create Audio TranscriptionTool to transcribe audio files to text via OpenAI Audio Transcriptions API.
Create Audio TranslationTool to translate audio files to English text via OpenAI Audio Translations API.
Create BatchTool to create and execute a batch from an uploaded file of requests.
Create Chat CompletionTool to create a chat completion response from OpenAI models.
Create Completion (Legacy)Tool to generate text completions using OpenAI's legacy Completions API.
Create ContainerTool to create a new container with configurable memory, expiration, file access, and network policies.
Create Container FileTool to create a file in a container.
Create ConversationTool to create a new conversation for multi-turn interactions.
Create Conversation ItemsTool to create items in a conversation with the given ID.
Create EmbeddingsTool to generate text embeddings via the OpenAI embeddings endpoint.
Create EvalTool to create an evaluation structure for testing a model's performance.
Create Evaluation RunTool to create a new evaluation run for testing model configurations.
Create fine-tuning jobTool to create a fine-tuning job which begins the process of creating a new model from a given dataset.
Generate ImageTool to generate an image via the OpenAI Images API and return hosted image asset URL and metadata.
Edit ImageTool to create edited or extended images via OpenAI Images Edit API.
Create Image VariationTool to create a variation of a given image using the OpenAI Images API.
Create MessageTool to create a new message in a specific thread.
Create ModerationTool to classify text and/or image inputs for potentially harmful content via the OpenAI Moderation API.
Create Realtime CallTool to create a Realtime API call over WebRTC and receive the SDP answer needed to complete the peer connection.
Create Realtime Client SecretTool to create an ephemeral client secret for authenticating Realtime API connections.
Create Realtime SessionTool to create an ephemeral API token for client-side Realtime API applications.
Create Realtime Transcription SessionTool to create an ephemeral API token for realtime transcriptions via the Realtime API.
Create ResponseTool to generate a one-shot model response via the Responses API.
Create RunTool to create a run on a thread with an assistant.
Create SkillTool to create a skill from uploaded files.
Create Speech (TTS)Tool to generate text-to-speech audio using OpenAI's Audio API.
Create ThreadTool to create a new thread.
Create Thread And RunTool to create a thread and run it in one request.
Create UploadTool to create an intermediate Upload object for large file uploads.
Create Vector StoreTool to create a new vector store.
Create Vector Store FileTool to create a vector store file by attaching a File to a vector store.
Create vector store file batchTool to create a vector store file batch.
Create VideoTool to create a video using Sora models via the OpenAI Videos API.
Create Video RemixTool to create a video remix from an existing generated video using OpenAI's Video API.
Delete assistantTool to delete a specific assistant by its ID.
Delete chat completionTool to delete a stored chat completion by its ID.
Delete containerTool to delete a specific container by its ID.
Delete container fileTool to delete a file from a container.
Delete conversationTool to delete a conversation by its ID.
Delete conversation itemTool to delete an item from a conversation with the given IDs.
Delete evaluationTool to delete a specific evaluation by its ID.
Delete evaluation runTool to delete an evaluation run.
Delete fileTool to delete a file by its ID after confirming the target.
Delete messageTool to delete a message from a thread.
Delete responseTool to delete a model response with the given ID.
Delete skillTool to delete a specific skill by its ID.
Delete threadTool to delete a thread by its ID.
Delete Vector StoreTool to delete a vector store.
Delete Vector Store FileTool to delete a vector store file.
Delete videoTool to delete a video by its ID.
Download fileTool to download the contents of a specified file by its ID.
Download Video ContentTool to download video content (MP4) or preview assets from OpenAI Videos API.
Get Chat CompletionTool to retrieve a stored chat completion.
Get Chat Completion MessagesTool to retrieve messages from a stored chat completion.
Get ChatKit threadTool to retrieve a ChatKit thread by its ID.
Get Conversation ItemTool to retrieve a single item from a conversation.
Get EvalTool to retrieve an evaluation by ID.
Get Evaluation RunTool to retrieve an evaluation run by ID to check status and results.
Get Eval Run Output ItemTool to retrieve a specific output item from an evaluation run by its ID.
Get eval run output itemsTool to get a list of output items for an evaluation run.
Get Evaluation RunsTool to get a paginated list of runs for an evaluation.
Get Input Token CountsTool to calculate input token counts for OpenAI API requests.
Get MessageTool to retrieve a specific message from a thread by its ID.
Get ResponseTool to retrieve a model response by ID.
Get Run StepTool to retrieve a specific run step from an Assistants API run to inspect detailed execution progress, view tool calls, or check message creation.
Get Vector StoreTool to retrieve a vector store by its ID.
Get Vector Store FileTool to retrieve a file from a vector store.
Get Vector Store File BatchTool to retrieve a vector store file batch.
Get VideoTool to retrieve a video generation job by its unique identifier.
List AssistantsTool to list assistants to discover the correct assistant_id by name or metadata.
List BatchesTool to list your organization's batches.
List Chat CompletionsTool to list stored chat completions that were created with the `store` parameter set to true.
List ChatKit thread itemsTool to list ChatKit thread items.
List container filesTool to list files in a container.
List ContainersTool to list containers.
List Conversation ItemsTool to list all items for a conversation with the given ID.
List enginesTool to list available engines and their basic information.
List EvalsTool to list evaluations for a project.
List filesTool to retrieve a list of files uploaded to your organization/project context.
List Files in Vector Store BatchTool to list vector store files in a batch.
List fine-tunesTool to list your organization's fine-tuning jobs.
List fine-tuning job eventsTool to get status updates for a fine-tuning job.
List fine-tuning job checkpointsTool to list checkpoints for a fine-tuning job.
List Input ItemsTool to retrieve input items for a given response from the OpenAI Responses API.
List MessagesTool to list messages in an Assistants thread to fetch the assistant's generated outputs after a run completes.
List modelsTool to list available models scoped to the current account/organization — some public models may be absent due to permissions.
List RunsTool to list runs belonging to a thread.
List Run StepsTool to list run steps for an Assistants API run to track detailed execution progress, inspect tool calls, and view message creation events.
List SkillsTool to list skills.
List ChatKit ThreadsTool to list ChatKit threads with pagination and filtering.
List Vector Store FilesTool to list files in a vector store.
List Vector StoresTool to list vector stores to discover available vector stores by name or metadata.
List VideosTool to list all video generation jobs.
Modify AssistantTool to modify an existing assistant.
Modify MessageTool to modify an existing message's metadata in a thread.
Modify RunTool to modify a run's metadata.
Modify threadTool to modify an existing thread's metadata.
Modify Vector StoreTool to modify an existing vector store.
Retrieve assistantTool to retrieve details of a specific assistant.
Retrieve BatchTool to retrieve a batch by ID to check its status, progress, and results.
Retrieve containerTool to retrieve details of a specific container by its ID.
Retrieve container fileTool to retrieve metadata for a specific file in a container.
Retrieve container file contentTool to retrieve the content of a file within a container.
Retrieve engineTool to retrieve details of a specific engine.
Retrieve fileTool to retrieve information about a specific file.
Retrieve fine-tuning jobTool to retrieve information about a fine-tuning job.
Retrieve modelTool to retrieve details of a specific model, confirming its metadata (ownership, created date) and verifying access under your org — a model appearing in OPENAI_LIST_MODELS does not guarantee access.
Retrieve runTool to retrieve an Assistants run by ID to check status, errors, and usage.
Retrieve threadTool to retrieve metadata of a specific thread by its ID — does not include message bodies or assistant replies (those require a completed run and separate message listing).
Retrieve Vector Store File ContentTool to retrieve the parsed contents of a vector store file.
Run graderTool to run a grader to evaluate model performance on a given sample.
Search Vector StoreTool to search a vector store for relevant chunks based on a query and file attributes filter.
Submit Tool Outputs to RunTool to submit tool call outputs to continue a run that requires action.
Update Chat CompletionTool to update metadata for a stored chat completion.
Update ConversationTool to update a conversation's metadata.
Update EvalTool to update certain properties of an evaluation (name and metadata).
Update Vector Store File AttributesTool to update custom attributes on a vector store file.
Upload fileTool to upload a file for use across OpenAI endpoints.
Validate grader configurationTool to validate a grader configuration for fine-tuning jobs.

What is the Composio tool router, and how does it fit here?

What is Composio SDK?

Composio's Composio SDK helps agents find the right tools for a task at runtime. You can plug in multiple toolkits (like Gmail, HubSpot, and GitHub), and the agent will identify the relevant app and action to complete multi-step workflows. This can reduce token usage and improve the reliability of tool calls. Read more here: Getting started with Composio SDK

The tool router generates a secure MCP URL that your agents can access to perform actions.

How the Composio SDK works

The Composio SDK follows a three-phase workflow:

  1. Discovery: Searches for tools matching your task and returns relevant toolkits with their details.
  2. Authentication: Checks for active connections. If missing, creates an auth config and returns a connection URL via Auth Link.
  3. Execution: Executes the action using the authenticated connection.

Step-by-step Guide

Prerequisites

Before you begin, make sure you have:
  • Node.js and npm installed
  • A Composio account with API key
  • An OpenAI API key

Getting API Keys for OpenAI and Composio

OpenAI API Key
  • Go to the OpenAI dashboard and create an API key. You'll need credits to use the models, or you can connect to another model provider.
  • Keep the API key safe.
Composio API Key
  • Log in to the Composio dashboard.
  • Navigate to your API settings and generate a new API key.
  • Store this key securely as you'll need it for authentication.

Install required dependencies

bash
npm install @ai-sdk/openai @ai-sdk/mcp @composio/core ai dotenv

First, install the necessary packages for your project.

What you're installing:

  • @ai-sdk/openai: Vercel AI SDK's OpenAI provider
  • @ai-sdk/mcp: MCP client for Vercel AI SDK
  • @composio/core: Composio SDK for tool integration
  • ai: Core Vercel AI SDK
  • dotenv: Environment variable management

Set up environment variables

bash
OPENAI_API_KEY=your_openai_api_key_here
COMPOSIO_API_KEY=your_composio_api_key_here
COMPOSIO_USER_ID=your_user_id_here

Create a .env file in your project root.

What's needed:

  • OPENAI_API_KEY: Your OpenAI API key for GPT model access
  • COMPOSIO_API_KEY: Your Composio API key for tool access
  • COMPOSIO_USER_ID: A unique identifier for the user session

Import required modules and validate environment

typescript
import "dotenv/config";
import { openai } from "@ai-sdk/openai";
import { Composio } from "@composio/core";
import * as readline from "readline";
import { streamText, type ModelMessage, stepCountIs } from "ai";
import { createMCPClient } from "@ai-sdk/mcp";

const composioAPIKey = process.env.COMPOSIO_API_KEY;
const composioUserID = process.env.COMPOSIO_USER_ID;

if (!process.env.OPENAI_API_KEY) throw new Error("OPENAI_API_KEY is not set");
if (!composioAPIKey) throw new Error("COMPOSIO_API_KEY is not set");
if (!composioUserID) throw new Error("COMPOSIO_USER_ID is not set");

const composio = new Composio({
  apiKey: composioAPIKey,
});
What's happening:
  • We're importing all necessary libraries including Vercel AI SDK's OpenAI provider and Composio
  • The dotenv/config import automatically loads environment variables
  • The MCP client import enables connection to Composio's tool server

Create Tool Router session and initialize MCP client

typescript
async function main() {
  // Create a tool router session for the user
  const session = await composio.create(composioUserID!, {
    toolkits: ["openai"],
  });

  const mcpUrl = session.mcp.url;
What's happening:
  • We're creating a Tool Router session that gives your agent access to Openai tools
  • The create method takes the user ID and specifies which toolkits should be available
  • The returned mcp object contains the URL and authentication headers needed to connect to the MCP server
  • This session provides access to all Openai-related tools through the MCP protocol

Connect to MCP server and retrieve tools

typescript
const mcpClient = await createMCPClient({
  transport: {
    type: "http",
    url: mcpUrl,
    headers: session.mcp.headers, // Authentication headers for the Composio MCP server
  },
});

const tools = await mcpClient.tools();
What's happening:
  • We're creating an MCP client that connects to our Composio Tool Router session via HTTP
  • The mcp.url provides the endpoint, and mcp.headers contains authentication credentials
  • The type: "http" is important - Composio requires HTTP transport
  • tools() retrieves all available Openai tools that the agent can use

Initialize conversation and CLI interface

typescript
let messages: ModelMessage[] = [];

console.log("Chat started! Type 'exit' or 'quit' to end the conversation.\n");
console.log(
  "Ask any questions related to openai, like summarize my last 5 emails, send an email, etc... :)))\n",
);

const rl = readline.createInterface({
  input: process.stdin,
  output: process.stdout,
  prompt: "> ",
});

rl.prompt();
What's happening:
  • We initialize an empty messages array to maintain conversation history
  • A readline interface is created to accept user input from the command line
  • Instructions are displayed to guide the user on how to interact with the agent

Handle user input and stream responses with real-time tool feedback

typescript
rl.on("line", async (userInput: string) => {
  const trimmedInput = userInput.trim();

  if (["exit", "quit", "bye"].includes(trimmedInput.toLowerCase())) {
    console.log("\nGoodbye!");
    rl.close();
    process.exit(0);
  }

  if (!trimmedInput) {
    rl.prompt();
    return;
  }

  messages.push({ role: "user", content: trimmedInput });
  console.log("\nAgent is thinking...\n");

  try {
    const stream = streamText({
      model: openai("gpt-5"),
      messages,
      tools,
      toolChoice: "auto",
      stopWhen: stepCountIs(10),
      onStepFinish: (step) => {
        for (const toolCall of step.toolCalls) {
          console.log(`[Using tool: ${toolCall.toolName}]`);
          }
          if (step.toolCalls.length > 0) {
            console.log(""); // Add space after tool calls
          }
        },
      });

      for await (const chunk of stream.textStream) {
        process.stdout.write(chunk);
      }

      console.log("\n\n---\n");

      // Get final result for message history
      const response = await stream.response;
      if (response?.messages?.length) {
        messages.push(...response.messages);
      }
    } catch (error) {
      console.error("\nAn error occurred while talking to the agent:");
      console.error(error);
      console.log(
        "\nYou can try again or restart the app if it keeps happening.\n",
      );
    } finally {
      rl.prompt();
    }
  });

  rl.on("close", async () => {
    await mcpClient.close();
    console.log("\n👋 Session ended.");
    process.exit(0);
  });
}

main().catch((err) => {
  console.error("Fatal error:", err);
  process.exit(1);
});
What's happening:
  • We use streamText instead of generateText to stream responses in real-time
  • toolChoice: "auto" allows the model to decide when to use Openai tools
  • stopWhen: stepCountIs(10) allows up to 10 steps for complex multi-tool operations
  • onStepFinish callback displays which tools are being used in real-time
  • We iterate through the text stream to create a typewriter effect as the agent responds
  • The complete response is added to conversation history to maintain context
  • Errors are caught and displayed with helpful retry suggestions

Complete Code

Here's the complete code to get you started with Openai and Vercel AI SDK:

typescript
import "dotenv/config";
import { openai } from "@ai-sdk/openai";
import { Composio } from "@composio/core";
import * as readline from "readline";
import { streamText, type ModelMessage, stepCountIs } from "ai";
import { createMCPClient } from "@ai-sdk/mcp";

const composioAPIKey = process.env.COMPOSIO_API_KEY;
const composioUserID = process.env.COMPOSIO_USER_ID;

if (!process.env.OPENAI_API_KEY) throw new Error("OPENAI_API_KEY is not set");
if (!composioAPIKey) throw new Error("COMPOSIO_API_KEY is not set");
if (!composioUserID) throw new Error("COMPOSIO_USER_ID is not set");

const composio = new Composio({
  apiKey: composioAPIKey,
});

async function main() {
  // Create a tool router session for the user
  const session = await composio.create(composioUserID!, {
    toolkits: ["openai"],
  });

  const mcpUrl = session.mcp.url;

  const mcpClient = await createMCPClient({
    transport: {
      type: "http",
      url: mcpUrl,
      headers: session.mcp.headers, // Authentication headers for the Composio MCP server
    },
  });

  const tools = await mcpClient.tools();

  let messages: ModelMessage[] = [];

  console.log("Chat started! Type 'exit' or 'quit' to end the conversation.\n");
  console.log(
    "Ask any questions related to openai, like summarize my last 5 emails, send an email, etc... :)))\n",
  );

  const rl = readline.createInterface({
    input: process.stdin,
    output: process.stdout,
    prompt: "> ",
  });

  rl.prompt();

  rl.on("line", async (userInput: string) => {
    const trimmedInput = userInput.trim();

    if (["exit", "quit", "bye"].includes(trimmedInput.toLowerCase())) {
      console.log("\nGoodbye!");
      rl.close();
      process.exit(0);
    }

    if (!trimmedInput) {
      rl.prompt();
      return;
    }

    messages.push({ role: "user", content: trimmedInput });
    console.log("\nAgent is thinking...\n");

    try {
      const stream = streamText({
        model: openai("gpt-5"),
        messages,
        tools,
        toolChoice: "auto",
        stopWhen: stepCountIs(10),
        onStepFinish: (step) => {
          for (const toolCall of step.toolCalls) {
            console.log(`[Using tool: ${toolCall.toolName}]`);
          }
          if (step.toolCalls.length > 0) {
            console.log(""); // Add space after tool calls
          }
        },
      });

      for await (const chunk of stream.textStream) {
        process.stdout.write(chunk);
      }

      console.log("\n\n---\n");

      // Get final result for message history
      const response = await stream.response;
      if (response?.messages?.length) {
        messages.push(...response.messages);
      }
    } catch (error) {
      console.error("\nAn error occurred while talking to the agent:");
      console.error(error);
      console.log(
        "\nYou can try again or restart the app if it keeps happening.\n",
      );
    } finally {
      rl.prompt();
    }
  });

  rl.on("close", async () => {
    await mcpClient.close();
    console.log("\n👋 Session ended.");
    process.exit(0);
  });
}

main().catch((err) => {
  console.error("Fatal error:", err);
  process.exit(1);
});

Conclusion

You've successfully built a Openai agent using the Vercel AI SDK with streaming capabilities! This implementation provides a powerful foundation for building AI applications with natural language interfaces and real-time feedback.

Key features of this implementation:

  • Real-time streaming responses for a better user experience with typewriter effect
  • Live tool execution feedback showing which tools are being used as the agent works
  • Dynamic tool loading through Composio's Tool Router with secure authentication
  • Multi-step tool execution with configurable step limits (up to 10 steps)
  • Comprehensive error handling for robust agent execution
  • Conversation history maintenance for context-aware responses

You can extend this further by adding custom error handling, implementing specific business logic, or integrating additional Composio toolkits to create multi-app workflows.

How to build Openai MCP Agent with another framework

FAQ

What are the differences in Tool Router MCP and Openai MCP?

With a standalone Openai MCP server, the agents and LLMs can only access a fixed set of Openai tools tied to that server. However, with the Composio Tool Router, agents can dynamically load tools from Openai and many other apps based on the task at hand, all through a single MCP endpoint.

Can I use Tool Router MCP with Vercel AI SDK v6?

Yes, you can. Vercel AI SDK v6 fully supports MCP integration. You get structured tool calling, message history handling, and model orchestration while Tool Router takes care of discovering and serving the right Openai tools.

Can I manage the permissions and scopes for Openai while using Tool Router?

Yes, absolutely. You can configure which Openai scopes and actions are allowed when connecting your account to Composio. You can also bring your own OAuth credentials or API configuration so you keep full control over what the agent can do.

How safe is my data with Composio Tool Router?

All sensitive data such as tokens, keys, and configuration is fully encrypted at rest and in transit. Composio is SOC 2 Type 2 compliant and follows strict security practices so your Openai data and credentials are handled as safely as possible.

Used by agents from

Context
Letta
glean
HubSpot
Agent.ai
Altera
DataStax
Entelligence
Rolai
Context
Letta
glean
HubSpot
Agent.ai
Altera
DataStax
Entelligence
Rolai
Context
Letta
glean
HubSpot
Agent.ai
Altera
DataStax
Entelligence
Rolai

Never worry about agent reliability

We handle tool reliability, observability, and security so you never have to second-guess an agent action.