Here is the process in short:

Compile the server code with webpack
Use target: "node" or target: "async-node"
Enabled HMR via --hot or HotModuleReplacementPlugin
Use webpack/hot/poll or webpack/hot/signal
The first polls the fs for updates (easy to use)
The second listens for a process event to check for updates (you need a way to send the signal)
Run the bundle with node.
You can't use existing HMR loaders like react-hot-loader or style-loader because they make no sense in a server environment. Just add manuall replacement code at the correct location (i. e. accept request handler like in the example)

You can't use the webpack-dev-server. It's a server which serves assets not a runner. Just run webpack --watch and node bundle.js. I would go the webpack/hot/poll?1000 route first. It's pretty easy to use and suitable for dev environments. For production (if you want to hot update your production server) the signal approach is better suited.

webpack = require 'webpack'

module.exports =
 entry: [
  'webpack/hot/poll?1000' # <-- 轮询更新内容的代码
  './src/main' # <-- 项目入口
 ]
 target: 'node' # <-- 指明编译方式为 node
 output:
  path: 'build/'
  filename: 'bundle.js' # <-- 编译结果的文件名
 module:
  loaders: [
   {test: /\.coffee/, loader: 'coffee'}
  ]
 plugins: [
  new webpack.HotModuleReplacementPlugin() # <-- 照常启动 hot mode
 ]
 resolve:
  extensions: ['.js', '', '.coffee']

npm i
webpack --watch & # <-- watch 模式
node build/bundle.js # <-- 运行的是打包结果的代码

exports.data = 'code 5'

exports.printSelf = ->
 console.log 'doing 3'

lib = require './lib'

console.log lib.data
lib.printSelf()

counter = 0
setInterval ->
 counter += 1
 console.log counter
, 2000

if module.hot
 module.hot.accept './lib', ->
  lib = require './lib'

  console.log lib.data
  lib.printSelf()