The Problem I Was Trying to Solve§

Standard Model Context Protocol (MCP) servers run over WebSockets or Server-Sent Events (SSE). While this is perfect for local desktop clients (like Cursor or Claude Desktop), it performs poorly in cloud environments. WebSockets and SSE maintain persistent connections, which break behind standard HTTP load balancers and make auto-scaling in serverless setups (like AWS Lambda or Vercel Functions) impossible. We needed to deploy a high-throughput, completely stateless MCP server using a Streamable HTTP architecture.

Tools and Setup (auto-link injection fires here)§

Our deployment pipeline utilized:

  • Next.js Route Handlers running on serverless runtimes.
  • Next.js Edge Runtime to stream responses chunk-by-chunk.
  • A custom transport client that maps stateless POST requests to standard MCP JSON-RPC protocol states.
// Next.js Edge Route Handler returning an MCP stream
export const runtime = 'edge';

export async function POST(req) {
  const encoder = new TextEncoder();
  const stream = new ReadableStream({
    async start(controller) {
      // Stream JSON-RPC chunks dynamically
      controller.enqueue(encoder.encode(JSON.stringify({ jsonrpc: "2.0", method: "tools/list" })));
      controller.close();
    }
  });
  return new Response(stream, { headers: { 'Content-Type': 'application/json-lines' } });
}

Step-by-Step: What I Actually Did§

1. Decoupling Connection State: We eliminated connection state from the MCP routing logic. Every request includes standard authentication headers, keeping the server stateless. 2. Streaming Chunk Format: We adopted JSON-lines (application/json-lines) over standard HTTP, allowing the agent client to process tool results as they stream in, rather than waiting for long database transactions to finish. 3. Load Balancer Configuration: We configured standard routing rules on our cloud gateway to load balance requests across serverless instances.

Results and Takeaways§

  • Scale to Zero: Serverless hosting allows the MCP server to scale down to zero when idle, saving significant infrastructure costs.
  • Low Latency: Streaming JSON-lines reduced time-to-first-token in multi-agent tool loops by 40%.
  • Design for Statelessness: When moving MCP servers from dev to cloud production, always prioritize stateless HTTP over persistent WebSockets.