diff --git a/hw/straylight/default.nix b/hw/straylight/default.nix
index 8330a3b..38d643a 100644
--- a/hw/straylight/default.nix
+++ b/hw/straylight/default.nix
@@ -200,6 +200,30 @@ in
       };
     };
 
+    # TLS for the llama.cpp API over the tailnet. Clients using
+    # https://straylight.scylla-hammerhead.ts.net/v1 hit Tailscale Serve on
+    # :443, which proxies to llama-server on localhost:11434. A boot oneshot
+    # is not enough: `tailscale serve reset` clears the config without
+    # stopping this unit, so a timer re-applies it.
+    systemd.services.tailscale-serve-llama = {
+      description = "Tailscale HTTPS proxy for llama-server";
+      after = [ "tailscaled.service" "llama-server.service" ];
+      wants = [ "tailscaled.service" ];
+      serviceConfig = {
+        Type = "oneshot";
+        ExecStart = "${pkgs.tailscale}/bin/tailscale serve --bg --https=443 http://127.0.0.1:11434";
+      };
+    };
+    systemd.timers.tailscale-serve-llama = {
+      description = "Re-apply Tailscale HTTPS proxy for llama-server";
+      wantedBy = [ "timers.target" ];
+      timerConfig = {
+        OnBootSec = "15s";
+        OnUnitActiveSec = "1min";
+        Unit = "tailscale-serve-llama.service";
+      };
+    };
+
     # Local ASR: OpenAI-compatible /v1/audio/transcriptions on 11435.
     # CPU-only so it does not contend with llama-server for GTT.
     users.users.whisper-server = {