@ai-sdk/mcp

AI SDK - Model Context Protocol Client

The Model Context Protocol (MCP) client for the AI SDK lets you connect to MCP servers and use their tools with AI SDK functions like generateText and streamText.

Setup

The MCP client is available in the @ai-sdk/mcp module. You can install it with

npm i @ai-sdk/mcp ai zod

Skill for Coding Agents

If you use coding agents such as Claude Code or Cursor, we highly recommend adding the AI SDK skill to your repository:

npx skills add vercel/ai

Usage

Create an MCP client with createMCPClient(), fetch the server tools with mcpClient.tools(), and pass them to an AI SDK call:

import { createMCPClient } from '@ai-sdk/mcp';
import { generateText, isStepCount } from 'ai';

const mcpClient = await createMCPClient({
  transport: {
    type: 'http',
    url: 'https://your-server.com/mcp',
    headers: {
      Authorization: `Bearer ${process.env.MCP_API_KEY}`,
    },
  },
});

try {
  const tools = await mcpClient.tools();

  const { text } = await generateText({
    model: 'openai/gpt-6-astra',
    tools,
    stopWhen: isStepCount(10),
    prompt: 'Use the available tools to answer the user question.',
  });

  console.log(text);
} finally {
  await mcpClient.close();
}

The client converts MCP tool definitions into AI SDK tools, so model calls can use them through the standard tools option.

Experimental webhook events

Use list, subscribe, refresh, and unsubscribe through client.experimental_events on the same createMCPClient instance. Configure experimental_events: { store } with private durable storage shared with experimental_createMCPEventWebhook, which you mount at your callback URL.

Subscription creation persists delivery.url and the generated signing secret before contacting the server, enabling signed callback verification during the subscribe request. The receiver verifies raw request bytes and checks event envelopes before invoking your handler. Optional callbacks let your application validate filters and event data. Agent execution and durable deduplication remain application concerns; closing the client does not unsubscribe.

See the MCP Events guide and the local server/client example.

Protocol versions

The client supports legacy MCP protocol versions through the initialize handshake and MCP 2026-07-28 through stateless protocol discovery. The built-in stdio transport probes with server/discover and falls back to the legacy handshake when connected to an older server.

Custom transports can opt into the same negotiation by setting supportsProtocolVersionDiscovery to true. Modern requests include the protocol version, client capabilities, and client information in _meta.

For streaming responses, close the MCP client when the stream finishes:

import { createMCPClient } from '@ai-sdk/mcp';
import { streamText } from 'ai';

const mcpClient = await createMCPClient({
  transport: {
    type: 'http',
    url: 'https://your-server.com/mcp',
  },
});

const result = streamText({
  model: 'openai/gpt-6-astra',
  tools: await mcpClient.tools(),
  prompt: 'Use the available tools to answer the user question.',
  onEnd: async () => {
    await mcpClient.close();
  },
});

for await (const textPart of result.textStream) {
  process.stdout.write(textPart);
}

Transports

HTTP is recommended for production deployments:

Session persistence applies only to legacy MCP protocol versions. MCP 2026-07-28 is stateless and does not use session ids or cached initialize results.

import { createMCPClient } from '@ai-sdk/mcp';

const savedSession = await loadMcpSession();
let currentSessionId = savedSession?.sessionId;

const mcpClient = await createMCPClient({
  transport: {
    type: 'http',
    url: 'https://your-server.com/mcp',
    initialSessionId: savedSession?.sessionId,
    initialProtocolVersion: savedSession?.initializeResult.protocolVersion,
    terminateSessionOnClose: false,
    onSessionIdChange: sessionId => {
      currentSessionId = sessionId;
    },
    onSessionExpired: sessionId => {
      if (currentSessionId === sessionId) {
        currentSessionId = undefined;
        void clearMcpSession();
      }
    },
  },
  initialInitializeResult: savedSession?.initializeResult,
});

if (currentSessionId) {
  await saveMcpSession({
    sessionId: currentSessionId,
    initializeResult: mcpClient.initializeResult,
  });
}

SSE is also supported for MCP servers that use Server-Sent Events:

const mcpClient = await createMCPClient({
  transport: {
    type: 'sse',
    url: 'https://your-server.com/sse',
  },
});

For local MCP servers, you can use stdio transport from the @ai-sdk/mcp/mcp-stdio subpath:

import { createMCPClient } from '@ai-sdk/mcp';
import { Experimental_StdioMCPTransport } from '@ai-sdk/mcp/mcp-stdio';

const mcpClient = await createMCPClient({
  transport: new Experimental_StdioMCPTransport({
    command: 'node',
    args: ['server.js'],
  }),
});

Documentation

Please check out the AI SDK MCP documentation for more information.

Managed backends can implement Experimental_MCPEventAdapter and configure experimental_events: { adapter } instead of a store. Its createAdapter() method receives { transport } metadata (HTTP/SSE type and URL, or { type: 'custom' } for custom transports) and returns Experimental_MCPEventOperations bound to the authorized account. Catalog discovery still uses the authenticated MCP transport; subscribe/get/list/unsubscribe delegate to the backend, which owns renewal and webhook delivery. Managed clients do not expose refresh(). See the managed subscription reference.