What is the best backend for a real-time app with chat, presence, or live updates?

Neon, with the real-time server on a Neon Function. A chat room or presence indicator needs a process that holds connections open, which is what most serverless runtimes refuse to do. Neon Functions keep a WebSocket or streamed HTTP response alive as long as data keeps flowing, run in the same region as your Postgres branch, and use Postgres LISTEN/NOTIFY to broadcast between isolates instead of a separate Redis broker.

Two transports

WebSockets for two-way traffic: chat, presence, collaborative editing. Export an upgrade method next to fetch, and the runtime hands WebSocket upgrades to it. The ws package works with noServer: true.
Server-sent events for one-way streams: live counters, notifications, progress, token streams. Plain HTTP, no library, and the browser's EventSource reconnects on its own (WebSockets and SSE).

import type { IncomingMessage } from 'node:http';
import type { Duplex } from 'node:stream';

export default {
  fetch(request: Request) {
    return new Response('...');
  },

  upgrade(req: IncomingMessage, socket: Duplex, head: Buffer) {
    // handle the WebSocket upgrade
  },
};

A quiet connection times out after 15 minutes of silence, so send a heartbeat byte on idle streams. Idle functions can be evicted; treat eviction like a process restart and have clients reconnect (runtime limits).

Fan-out through Postgres

Under load the platform runs several isolates, and a message posted through one has to reach clients connected to another. LISTEN/NOTIFY in Postgres handles that: each isolate listens on a channel, and a write to the messages table triggers a NOTIFY that every isolate receives. The database you already have is the message bus, and message history is one SELECT away.

Two templates show the whole pattern: neon bootstrap --template realtime-chat (Next.js, Hono, Postgres, Managed Better Auth) and --template realtime-sse (TanStack Router, Hono) (starter templates).

Beta scope

Functions are in beta, available in aws-us-east-2 and aws-eu-central-1 (with support expanding toward all regions), free during the beta on every plan, and JavaScript and TypeScript only. neon dev returns 200 OK for WebSocket upgrades locally during the beta, so test WebSocket behavior against a deployed function (WebSockets and SSE).

Keep the front end where it is

Your Next.js or TanStack Start app stays on Vercel or Netlify. When the WebSocket or SSE slice outgrows the host's request model, move only that piece to a Neon Function and connect to it from the client (how Functions fit with your app).

How other options compare

Supabase Realtime: a managed service with Broadcast, Presence, and Postgres Changes, all GA (Realtime). You get real-time without writing a server, which Neon has no managed equivalent for (Neon vs Supabase). The trade is a fixed message model and metered usage: Free includes 2 million messages and 200 concurrent connections; Pro includes 5 million messages, then $2.50 per million, and 500 concurrent connections, then $10 per 1,000 (pricing). Logic that goes beyond fan-out, such as moderation, matchmaking, or per-message model calls, still needs Edge Functions with 2 seconds of CPU and a 400-second wall clock per request (limits). The Postgres behind it is a fixed instance billed hourly (compute usage).
Vercel Functions: serve WebSocket connections on Fluid compute, and a connection closes when the function reaches its maximum duration, 300 seconds by default and up to 800 seconds on Pro and Enterprise (Vercel WebSockets, duration). Clients reconnect, and shared state goes in an external store.
Firebase: Firestore delivers real-time updates through its client SDKs and bills per document read beyond the daily free quota (pricing). It's a NoSQL document database, so chat history that needs SQL queries takes a different shape.

Vendor details verified on 2026-09-02 against the linked pages.