diff --git a/backend/app/controllers/admin/stream_nodes_controller.rb b/backend/app/controllers/admin/stream_nodes_controller.rb
index 3a492ba..46e0283 100644
--- a/backend/app/controllers/admin/stream_nodes_controller.rb
+++ b/backend/app/controllers/admin/stream_nodes_controller.rb
@@ -9,6 +9,7 @@ module Admin
       @cloud_provider = ENV.fetch("STREAM_CLOUD_PROVIDER", "local_lab")
       @lab_hosts = lab_dns_snippet
       @hetzner_configured = ENV["HCLOUD_TOKEN"].present?
+      @autoscale_metrics = Streams::Autoscaler.metrics
     end
 
     def create
diff --git a/backend/app/jobs/streams/autoscaler_job.rb b/backend/app/jobs/streams/autoscaler_job.rb
new file mode 100644
index 0000000..a0e44c4
--- /dev/null
+++ b/backend/app/jobs/streams/autoscaler_job.rb
@@ -0,0 +1,33 @@
+# frozen_string_literal: true
+
+module Streams
+  class AutoscalerJob
+    include Sidekiq::Job
+
+    sidekiq_options retry: 1, queue: "default"
+
+    INTERVAL_SECS = ENV.fetch("STREAM_AUTOSCALE_INTERVAL_SECS", "60").to_i
+    REDIS_CHAIN_KEY = "streams:autoscaler:chain"
+
+    def self.ensure_chain
+      return unless redis
+      return if redis.get(REDIS_CHAIN_KEY)
+
+      redis.set(REDIS_CHAIN_KEY, "1", ex: INTERVAL_SECS * 2)
+      perform_in(INTERVAL_SECS)
+    end
+
+    def self.redis
+      @redis ||= Redis.new(url: ENV.fetch("REDIS_URL", "redis://localhost:6379/0"))
+    rescue Redis::CannotConnectError
+      nil
+    end
+
+    def perform
+      Streams::Autoscaler.reconcile!
+    ensure
+      self.class.redis&.set(REDIS_CHAIN_KEY, "1", ex: INTERVAL_SECS * 2)
+      self.class.perform_in(INTERVAL_SECS)
+    end
+  end
+end
diff --git a/backend/app/services/streams/autoscaler.rb b/backend/app/services/streams/autoscaler.rb
new file mode 100644
index 0000000..07f6d62
--- /dev/null
+++ b/backend/app/services/streams/autoscaler.rb
@@ -0,0 +1,158 @@
+# frozen_string_literal: true
+
+module Streams
+  # Scale-out / warm spare / scale-in dei nodi overflow (lab o Hetzner).
+  # Kill-switch: STREAM_AUTOSCALE_ENABLED!=1 → no-op.
+  class Autoscaler
+    Result = Struct.new(:actions, :metrics, :skipped, :error, keyword_init: true)
+
+    LOCK_KEY = "streams:autoscaler:lock"
+
+    class << self
+      def enabled?
+        ENV["STREAM_AUTOSCALE_ENABLED"] == "1"
+      end
+
+      def soft_free_slots
+        ENV.fetch("STREAM_AUTOSCALE_SOFT_FREE_SLOTS", "2").to_i
+      end
+
+      def warm_spare_min
+        ENV.fetch("STREAM_AUTOSCALE_WARM_SPARE", "1").to_i
+      end
+
+      def idle_minutes
+        ENV.fetch("STREAM_AUTOSCALE_IDLE_MINUTES", "30").to_i
+      end
+
+      def max_overflow_nodes
+        ENV.fetch("STREAM_AUTOSCALE_MAX_NODES", "5").to_i
+      end
+
+      def kind
+        ENV.fetch("STREAM_AUTOSCALE_KIND", "lab") # lab|cloud
+      end
+
+      def reconcile!(provisioner: nil)
+        return Result.new(skipped: true, actions: [], metrics: metrics) unless enabled?
+
+        redis = Redis.new(url: ENV.fetch("REDIS_URL", "redis://localhost:6379/0"))
+        unless redis.set(LOCK_KEY, worker_id, nx: true, ex: 55)
+          return Result.new(skipped: true, actions: [], metrics: metrics, error: "locked")
+        end
+
+        begin
+          new(provisioner: provisioner).reconcile!
+        ensure
+          redis.del(LOCK_KEY)
+        end
+      end
+
+      def metrics
+        NodeRegistry.ensure_home_from_env!
+        nodes = StreamNode.ready.to_a
+        overflow = StreamNode.where.not(slug: NodeRegistry::HOME_SLUG)
+          .where(status: %w[ready draining provisioning]).to_a
+        {
+          free_slots: nodes.sum(&:free_slots),
+          ready_nodes: nodes.size,
+          spare_ready: nodes.count { |n| n.slug != NodeRegistry::HOME_SLUG && n.active_publishers.zero? },
+          overflow_nodes: overflow.size,
+          soft_free_slots: soft_free_slots,
+          warm_spare_min: warm_spare_min,
+          max_overflow_nodes: max_overflow_nodes,
+          enabled: enabled?,
+          kind: kind
+        }
+      end
+
+      def worker_id
+        ENV.fetch("HOSTNAME", "autoscaler")
+      end
+    end
+
+    def initialize(provisioner: nil)
+      @provisioner = provisioner || NodeProvisioner.new
+    end
+
+    def reconcile!
+      actions = []
+      m = self.class.metrics
+
+      if need_capacity?(m) && can_provision?(m)
+        provision_overflow!
+        actions << :scale_out
+        m = self.class.metrics
+      end
+
+      if warm_spare_desired?(m) && m[:spare_ready] < self.class.warm_spare_min && can_provision?(m)
+        provision_overflow!
+        actions << :warm_spare
+        m = self.class.metrics
+      end
+
+      scale_in_candidates.each do |node|
+        next if keep_as_warm_spare?(node)
+
+        @provisioner.decommission!(node)
+        actions << :"scale_in_#{node.slug}"
+        m = self.class.metrics
+      rescue NodeProvisioner::BusyError, NodeProvisioner::Error => e
+        Rails.logger.warn("[Streams::Autoscaler] scale-in #{node.slug}: #{e.message}")
+      end
+
+      Rails.logger.info("[Streams::Autoscaler] actions=#{actions.inspect} metrics=#{m.inspect}")
+      Result.new(actions: actions, metrics: m, skipped: false)
+    end
+
+    private
+
+    def need_capacity?(m)
+      m[:free_slots] <= self.class.soft_free_slots
+    end
+
+    def warm_spare_desired?(m)
+      return false if self.class.warm_spare_min <= 0
+
+      need_capacity?(m) || overflow_in_use?
+    end
+
+    def overflow_in_use?
+      StreamNode.ready.where.not(slug: NodeRegistry::HOME_SLUG).any? { |n| n.active_publishers.positive? }
+    end
+
+    def can_provision?(m)
+      m[:overflow_nodes] < self.class.max_overflow_nodes
+    end
+
+    def provision_overflow!
+      case self.class.kind
+      when "cloud" then @provisioner.provision_cloud!
+      else @provisioner.provision_lab!
+      end
+    end
+
+    def scale_in_candidates
+      StreamNode.where.not(slug: NodeRegistry::HOME_SLUG)
+        .where(status: %w[ready draining])
+        .order(:created_at)
+        .select { |n| n.active_publishers.zero? && idle_long_enough?(n) }
+    end
+
+    def idle_long_enough?(node)
+      idle_since(node) <= self.class.idle_minutes.minutes.ago
+    end
+
+    def idle_since(node)
+      last_end = node.stream_sessions.where(status: %w[ended error]).maximum(:ended_at)
+      last_end || node.created_at
+    end
+
+    def keep_as_warm_spare?(node)
+      return false unless warm_spare_desired?(self.class.metrics)
+
+      spares = StreamNode.ready.where.not(slug: NodeRegistry::HOME_SLUG).select { |n| n.active_publishers.zero? }
+      spares.size <= self.class.warm_spare_min && spares.map(&:id).include?(node.id)
+    end
+  end
+end
diff --git a/backend/app/views/admin/stream_nodes/index.html.erb b/backend/app/views/admin/stream_nodes/index.html.erb
index dd8f2a2..d45b759 100644
--- a/backend/app/views/admin/stream_nodes/index.html.erb
+++ b/backend/app/views/admin/stream_nodes/index.html.erb
@@ -4,6 +4,20 @@
   <%= t("admin.stream_nodes.providers", cloud: @cloud_provider, dns: @dns_provider) %>
 </p>
 
+<p class="admin-muted">
+  <%= t(
+        "admin.stream_nodes.autoscale",
+        enabled: (@autoscale_metrics[:enabled] ? "ON" : "OFF"),
+        free: @autoscale_metrics[:free_slots],
+        soft: @autoscale_metrics[:soft_free_slots],
+        spare: @autoscale_metrics[:spare_ready],
+        warm: @autoscale_metrics[:warm_spare_min],
+        overflow: @autoscale_metrics[:overflow_nodes],
+        max: @autoscale_metrics[:max_overflow_nodes],
+        kind: @autoscale_metrics[:kind]
+      ) %>
+</p>
+
 <p>
   <%= button_to t("admin.stream_nodes.provision_lab"), admin_stream_nodes_path, method: :post, params: { kind: "lab" }, class: "admin-btn" %>
   <% if @hetzner_configured %>
diff --git a/backend/config/initializers/sidekiq.rb b/backend/config/initializers/sidekiq.rb
index 75e9ccf..307d522 100644
--- a/backend/config/initializers/sidekiq.rb
+++ b/backend/config/initializers/sidekiq.rb
@@ -6,6 +6,7 @@ Sidekiq.configure_server do |config|
   config.on(:startup) do
     StreamPublisherSyncJob.ensure_chain
     Ops::HealthMonitorJob.ensure_chain
+    Streams::AutoscalerJob.ensure_chain
   end
 end
 
diff --git a/backend/config/locales/admin.en.yml b/backend/config/locales/admin.en.yml
index 4c1da92..ded5346 100644
--- a/backend/config/locales/admin.en.yml
+++ b/backend/config/locales/admin.en.yml
@@ -118,6 +118,7 @@ en:
     stream_nodes:
       title: Streaming nodes
       providers: "Cloud provider: %{cloud} · DNS provider: %{dns}"
+      autoscale: "Autoscaler %{enabled} · free_slots=%{free} (soft≤%{soft}) · spare=%{spare}/%{warm} · overflow=%{overflow}/%{max} · kind=%{kind}"
       provision_lab: Provision lab node
       provision_cloud: Provision Hetzner node
       provision_cloud_confirm: "Create a Hetzner Cloud server + DNS record on mltv-stream.net? Billing applies until destroyed."
diff --git a/backend/config/locales/admin.it.yml b/backend/config/locales/admin.it.yml
index 37fc6ed..1b03118 100644
--- a/backend/config/locales/admin.it.yml
+++ b/backend/config/locales/admin.it.yml
@@ -118,6 +118,7 @@ it:
     stream_nodes:
       title: Nodi streaming
       providers: "Cloud provider: %{cloud} · DNS provider: %{dns}"
+      autoscale: "Autoscaler %{enabled} · free_slots=%{free} (soft≤%{soft}) · spare=%{spare}/%{warm} · overflow=%{overflow}/%{max} · kind=%{kind}"
       provision_lab: Provisiona nodo lab
       provision_cloud: Provisiona nodo Hetzner
       provision_cloud_confirm: "Creare un server Hetzner Cloud + record DNS su mltv-stream.net? Verrà addebitato fino allo spegnimento."
diff --git a/backend/spec/services/streams/autoscaler_spec.rb b/backend/spec/services/streams/autoscaler_spec.rb
new file mode 100644
index 0000000..c795652
--- /dev/null
+++ b/backend/spec/services/streams/autoscaler_spec.rb
@@ -0,0 +1,125 @@
+# frozen_string_literal: true
+
+require "rails_helper"
+
+RSpec.describe Streams::Autoscaler do
+  def with_env(vars)
+    previous = vars.keys.index_with { |k| ENV[k] }
+    vars.each { |k, v| ENV[k] = v }
+    yield
+  ensure
+    previous.each { |k, v| v.nil? ? ENV.delete(k) : ENV[k] = v }
+  end
+
+  let(:redis) { Redis.new(url: ENV.fetch("REDIS_URL", "redis://redis:6379/0")) }
+
+  before do
+    redis.del(Streams::Autoscaler::LOCK_KEY)
+    redis.del(Streams::DnsProviders::Lab::REDIS_KEY)
+  end
+
+  it "is a no-op when disabled" do
+    with_env("STREAM_AUTOSCALE_ENABLED" => "0") do
+      result = described_class.reconcile!
+      expect(result.skipped).to eq(true)
+      expect(result.actions).to eq([])
+    end
+  end
+
+  it "scales out and creates a warm spare when free slots are low" do
+    with_env(
+      "STREAM_AUTOSCALE_ENABLED" => "1",
+      "STREAM_AUTOSCALE_SOFT_FREE_SLOTS" => "2",
+      "STREAM_AUTOSCALE_WARM_SPARE" => "1",
+      "STREAM_AUTOSCALE_KIND" => "lab",
+      "STREAM_AUTOSCALE_MAX_NODES" => "5",
+      "STREAM_CLOUD_PROVIDER" => "local_lab",
+      "STREAM_DNS_PROVIDER" => "lab",
+      "STREAM_NODE_HOME_MAX_PUBLISHERS" => "2",
+      "MEDIAMTX_API_URL" => "http://mtx-home:9997",
+      "MEDIAMTX_RTMP_URL" => "rtmp://home.example:1935",
+      "HLS_PUBLIC_URL" => "https://home.example/hls"
+    ) do
+      home = Streams::NodeRegistry.ensure_home_from_env!
+      user = User.create!(email: "as@example.com", name: "A", password: "Password123", role: "coach")
+      club = Club.create!(name: "AS", sport: "volleyball")
+      team = club.teams.create!(name: "T", sport: "volleyball", slug: "as-t")
+      match = team.matches.create!(opponent_name: "X", scheduled_at: 1.hour.from_now)
+      # Fill home (2/2) → free_slots=0
+      2.times do
+        StreamSession.create!(match: match, user: user, platform: "matchlivetv", status: "live", stream_node: home)
+      end
+
+      provisioner = instance_double(Streams::NodeProvisioner)
+      created = []
+      allow(provisioner).to receive(:provision_lab!) do
+        node = StreamNode.create!(
+          slug: "ingest-lab-#{created.size + 1}",
+          hostname: "h#{created.size}.lab",
+          role: "lab",
+          status: "ready",
+          provider: "local",
+          rtmp_base_url: "rtmp://h:1935",
+          hls_base_url: "https://h/hls",
+          api_base_url: "http://h:9997",
+          max_publishers: 2,
+          max_relays: 2
+        )
+        created << node
+        node
+      end
+
+      result = described_class.reconcile!(provisioner: provisioner)
+      expect(result.actions).to include(:scale_out)
+      expect(created.size).to eq(1)
+      expect(result.metrics[:free_slots]).to be >= 2
+
+      # Riempi il nodo overflow → spare_ready=0 ma carico presente → warm spare
+      StreamSession.create!(
+        match: match, user: user, platform: "matchlivetv", status: "live", stream_node: created.first
+      )
+      StreamSession.create!(
+        match: match, user: user, platform: "matchlivetv", status: "live", stream_node: created.first
+      )
+
+      result2 = described_class.reconcile!(provisioner: provisioner)
+      expect(result2.actions).to include(:warm_spare).or include(:scale_out)
+      expect(created.size).to be >= 2
+    end
+  end
+
+  it "scales in an idle overflow node when warm spare is not required" do
+    with_env(
+      "STREAM_AUTOSCALE_ENABLED" => "1",
+      "STREAM_AUTOSCALE_SOFT_FREE_SLOTS" => "0",
+      "STREAM_AUTOSCALE_WARM_SPARE" => "0",
+      "STREAM_AUTOSCALE_IDLE_MINUTES" => "0",
+      "STREAM_NODE_HOME_MAX_PUBLISHERS" => "6",
+      "MEDIAMTX_API_URL" => "http://mtx-home:9997",
+      "MEDIAMTX_RTMP_URL" => "rtmp://home.example:1935",
+      "HLS_PUBLIC_URL" => "https://home.example/hls"
+    ) do
+      Streams::NodeRegistry.ensure_home_from_env!
+      idle = StreamNode.create!(
+        slug: "ingest-lab-99",
+        hostname: "idle.lab",
+        role: "lab",
+        status: "ready",
+        provider: "local",
+        provider_instance_id: "sim-x",
+        rtmp_base_url: "rtmp://h:1935",
+        hls_base_url: "https://h/hls",
+        api_base_url: "http://h:9997",
+        max_publishers: 2,
+        max_relays: 2,
+        created_at: 2.hours.ago
+      )
+
+      provisioner = instance_double(Streams::NodeProvisioner)
+      expect(provisioner).to receive(:decommission!).with(idle)
+
+      result = described_class.reconcile!(provisioner: provisioner)
+      expect(result.actions.map(&:to_s)).to include("scale_in_ingest-lab-99")
+    end
+  end
+end
diff --git a/docs/infrastructure/STREAMING_AUTOSCALE.md b/docs/infrastructure/STREAMING_AUTOSCALE.md
index 37badc9..fe14507 100644
--- a/docs/infrastructure/STREAMING_AUTOSCALE.md
+++ b/docs/infrastructure/STREAMING_AUTOSCALE.md
@@ -210,7 +210,7 @@ Cosa il lab **non** replica al 100%: tempi boot Hetzner, vSwitch, RTMP 4G multi-
 | **0 — Multi-nodo ready** | Registry + URL RTMP/HLS in API (`StreamNode`, `Streams::NodeRegistry`) | **implementata sul branch** |
 | **1 — Hetzner Cloud + DNS** | `HetznerCloudProvider` + `HetznerDnsProvider`, `mltv-stream.net`, cloud-init, WireGuard (§16) | **implementata sul branch** (WG ops manuale) |
 | **2 — Routing relay** | Coda `youtube_relay`, sticky owner, cap `RELAY_MAX_CONCURRENT` | **implementata sul branch** |
-| **3 — Autoscaler** | Soglie + warm spare (+1) | Automatico |
+| **3 — Autoscaler** | Soglie + warm spare (+1), kill-switch `STREAM_AUTOSCALE_ENABLED` | **implementata sul branch** |
 | **4 — Hardening** | Drain, budget, runbook, kill-switch | Picchi weekend |
 | **A — Auction (futuro)** | Migrazione control plane + vSwitch al posto di WireGuard | Hardware dedicato Hetzner |
 | **B2 (opz.)** | Switch object storage | Quando misurato |
@@ -250,6 +250,16 @@ Cosa il lab **non** replica al 100%: tempi boot Hetzner, vSwitch, RTMP 4G multi-
 - Cap per worker: `RELAY_MAX_CONCURRENT` (default 4) + set Redis `youtube_relay:owned:HOSTNAME`
 - Ensure a capacità piena → requeue 5s invece di avviare un secondo ffmpeg locale
 
+### Fase 3 — dettagli implementati
+
+- `Streams::Autoscaler` + job chain `Streams::AutoscalerJob` (come health monitor)
+- Kill-switch: `STREAM_AUTOSCALE_ENABLED=1` per attivare
+- Scale-out se `free_slots ≤ STREAM_AUTOSCALE_SOFT_FREE_SLOTS`
+- Warm spare (+N) se carico/soglia e `spare_ready < WARM_SPARE`
+- Scale-in overflow idle da `IDLE_MINUTES` (non tocca home; rispetta warm spare)
+- Cap `STREAM_AUTOSCALE_MAX_NODES`; kind `lab|cloud`
+- Metriche in admin Nodi stream
+
 Ordine di lavoro consigliato sul branch: **0 → L → 1 → 2 → 3 → 4**; **A** quando si decide di lasciare il Proxmox casa.
 
 ---
diff --git a/infra/.env.production.example b/infra/.env.production.example
index 1198f0e..bc61072 100644
--- a/infra/.env.production.example
+++ b/infra/.env.production.example
@@ -123,3 +123,12 @@ STREAM_NODE_ENV=prod
 STREAM_CLOUD_PROVIDER=local_lab
 RELAY_MAX_CONCURRENT=4
 YOUTUBE_RELAY_WORKER=1
+
+# Autoscaler (kill-switch: 0 finché WireGuard/smoke Cloud non sono OK)
+STREAM_AUTOSCALE_ENABLED=0
+STREAM_AUTOSCALE_KIND=lab
+STREAM_AUTOSCALE_SOFT_FREE_SLOTS=2
+STREAM_AUTOSCALE_WARM_SPARE=1
+STREAM_AUTOSCALE_IDLE_MINUTES=30
+STREAM_AUTOSCALE_MAX_NODES=5
+STREAM_AUTOSCALE_INTERVAL_SECS=60