AIHeroRight now, if you want to use an MCP server with the AI SDK, you need to run it locally on your machine. That means executing code from someone else's library with npx - which can feel risky.
But MCP servers are being deployed all over the world. So how do you contact them without running code locally?
The answer is HTTP transport. Instead of using standard IO transport to run an MCP server locally, you can point your AI SDK client at an external API endpoint.
The AI SDK supports three different transport types for connecting to MCP servers:
| Transport | Use Case | Trade-offs |
|---|---|---|
| Standard IO | Local MCP servers | Simpler setup, but executes code on your machine |
| HTTP | Remote MCP servers | No local code execution, but depends on external service |
| Server-Sent Events (SSE) | Real-time remote servers | Lower latency, but less widely supported |
The current code uses standard IO transport to run an MCP server locally. You'll need to understand what createMCPClient does and how it currently works.
Change the transport property in the createMCPClient call to use HTTP instead. You'll need to:
type to 'http'url pointing to the remote MCP serverheaders for authenticationRun pnpm run dev and test the same request: "Give me the latest issues on the mattpocock/ts-reset repo."
The agent should fetch data from the remote MCP server through HTTP instead of running code locally.
Notice how the HTTP transport:
Makes the setup simpler (no local process to manage)
Demands less from your server
Still gets the same results as the local setup
Think about what happens if the external service goes down. With HTTP transport, your entire agent depends on that external API being available. This is the main caveat of using remote MCP servers.
import { experimental_createMCPClient as createMCPClient } from '@ai-sdk/mcp';const mcpClient = await createMCPClient({transport: {type: 'http',url: 'https://api.githubcopilot.com/mcp',headers: {Authorization: `Bearer ${process.env.GITHUB_PERSONAL_ACCESS_TOKEN}`,},},});