diff --git a/hosts/mainframe/bin/stream b/hosts/mainframe/bin/stream
new file mode 100755
index 0000000..6e7ddac
--- /dev/null
+++ b/hosts/mainframe/bin/stream
@@ -0,0 +1,63 @@
+#!/usr/bin/env bash
+# Start and stop the Moonlight streaming stack on demand.
+#
+# It is not left running because it costs ~220 MiB of VRAM on the 3060 (sway
+# ~112, Sunshine ~108) and that card also holds the local LLM. With an 11.3 GB
+# model resident on a 12 GB card, 220 MiB is roughly a fifth of the remaining
+# headroom — enough to matter for KV cache growth.
+set -euo pipefail
+
+UNIT=sway-headless.service
+IDLE_MINUTES="${STREAM_IDLE_MINUTES:-15}"
+STAMP="${XDG_RUNTIME_DIR:-/tmp}/stream.lastseen"
+
+# A Moonlight client holds connections to the control and RTSP ports.
+clients_connected() {
+  ss -Htn state established '( sport = :47984 or sport = :47989 or sport = :48010 )' 2>/dev/null | grep -q .
+}
+
+vram() {
+  local used total
+  read -r used total < <(nvidia-smi --query-gpu=memory.used,memory.total \
+    --format=csv,noheader,nounits 2>/dev/null | tr -d ',')
+  printf '%s MiB used of %s, %s free' "$used" "$total" "$((total - used))"
+}
+
+case "${1:-status}" in
+  on|start)
+    systemctl --user start "$UNIT"
+    sleep 6
+    date +%s > "$STAMP"
+    printf 'streaming up — pair or connect at https://%s:47990\n' "$(uname -n)"
+    printf 'GPU: %s\n' "$(vram)"
+    ;;
+  off|stop)
+    systemctl --user stop "$UNIT"
+    sleep 3
+    rm -f "$STAMP"
+    printf 'streaming down\nGPU: %s\n' "$(vram)"
+    ;;
+  status)
+    printf 'compositor: %s\nsunshine:   %s\nclients:    %s\nGPU:        %s\n' \
+      "$(systemctl --user is-active "$UNIT")" \
+      "$(systemctl --user is-active sunshine.service)" \
+      "$(clients_connected && echo connected || echo none)" \
+      "$(vram)"
+    ;;
+  --idle-check)
+    # Called from sunshine-idle-stop.timer. Releases the GPU when nobody has
+    # been connected for IDLE_MINUTES, so a forgotten session does not sit on
+    # VRAM the inference stack wants.
+    systemctl --user is-active --quiet "$UNIT" || exit 0
+    if clients_connected; then date +%s > "$STAMP"; exit 0; fi
+    last=$(cat "$STAMP" 2>/dev/null || echo 0)
+    if (( last == 0 )); then date +%s > "$STAMP"; exit 0; fi
+    if (( ($(date +%s) - last) >= IDLE_MINUTES * 60 )); then
+      echo "no Moonlight client for ${IDLE_MINUTES}m; stopping to free the GPU"
+      systemctl --user stop "$UNIT"
+      rm -f "$STAMP"
+    fi
+    ;;
+  *)
+    echo "usage: stream [on|off|status]" >&2; exit 2 ;;
+esac
diff --git a/hosts/mainframe/systemd/sunshine-idle-stop.service b/hosts/mainframe/systemd/sunshine-idle-stop.service
new file mode 100644
index 0000000..a95de32
--- /dev/null
+++ b/hosts/mainframe/systemd/sunshine-idle-stop.service
@@ -0,0 +1,7 @@
+[Unit]
+Description=Release the GPU when no Moonlight client is connected
+ConditionHost=mainframe
+
+[Service]
+Type=oneshot
+ExecStart=%h/.dotfiles/hosts/mainframe/bin/stream --idle-check
diff --git a/hosts/mainframe/systemd/sunshine-idle-stop.timer b/hosts/mainframe/systemd/sunshine-idle-stop.timer
new file mode 100644
index 0000000..a096da0
--- /dev/null
+++ b/hosts/mainframe/systemd/sunshine-idle-stop.timer
@@ -0,0 +1,11 @@
+[Unit]
+Description=Check every 5 minutes whether the streaming stack can be shut down
+ConditionHost=mainframe
+
+[Timer]
+OnBootSec=5min
+OnUnitActiveSec=5min
+AccuracySec=30s
+
+[Install]
+WantedBy=timers.target
diff --git a/hosts/mainframe/systemd/sway-headless.service b/hosts/mainframe/systemd/sway-headless.service
index 6ea0b31..25f9beb 100644
--- a/hosts/mainframe/systemd/sway-headless.service
+++ b/hosts/mainframe/systemd/sway-headless.service
@@ -18,5 +18,6 @@ ExecStopPost=/usr/bin/systemctl --user stop sunshine.service
 Restart=on-failure
 RestartSec=5s
 
-[Install]
-WantedBy=default.target
+# Deliberately NOT WantedBy=default.target. The 3060 also holds the local LLM,
+# and this stack costs ~220 MiB of VRAM. Start it with `stream on`; the
+# idle-stop timer puts it away again.
diff --git a/meta/hosts/mainframe.conf.yaml b/meta/hosts/mainframe.conf.yaml
index 38b1b95..8a2e7d2 100644
--- a/meta/hosts/mainframe.conf.yaml
+++ b/meta/hosts/mainframe.conf.yaml
@@ -12,6 +12,8 @@
 - link:
     ~/.config/sway/headless.conf:                  hosts/mainframe/sway/headless.conf
     ~/.config/systemd/user/sway-headless.service:  hosts/mainframe/systemd/sway-headless.service
+    ~/.config/systemd/user/sunshine-idle-stop.service: hosts/mainframe/systemd/sunshine-idle-stop.service
+    ~/.config/systemd/user/sunshine-idle-stop.timer:   hosts/mainframe/systemd/sunshine-idle-stop.timer
     ~/.config/systemd/user/sunshine.service:       hosts/mainframe/systemd/sunshine.service
     ~/.config/inbox-tidy:            hosts/mainframe/config/inbox-tidy
     ~/.config/llmfit:                hosts/mainframe/config/llmfit
@@ -41,5 +43,5 @@
       description: Reloading user units
     - command: systemctl --user enable --now inbox-tidy.timer movie-recs.timer rent-utilities.timer ddns.timer
       description: Enabling the singleton timers (mainframe only)
-    - command: systemctl --user enable sway-headless.service sunshine.service
-      description: Enabling the headless compositor and Sunshine (not started here — it takes the GPU)
+    - command: systemctl --user enable --now sunshine-idle-stop.timer
+      description: Enabling the streaming idle-stop watchdog (the stack itself starts on demand via `stream on`)