diff --git a/ATHENA.md b/ATHENA.md
index 8a4d9fc..3d980be 100644
--- a/ATHENA.md
+++ b/ATHENA.md
@@ -13,9 +13,9 @@ Fehlersuche sie benötigt.
 - Lokale Konfiguration und Secrets: `/etc/mike-ai` (niemals in Git).
 - Benutzerzugriff auf KI-Dienste: über WireGuard, nicht über das Uni-LAN.
 - OpenAI-kompatible Modell-API: Profile Router auf Port 8081.
-- Oberfläche auf Athena: OpenWebUI. Der offizielle Hermes Agent läuft auf
-  Unraid unter `/mnt/nvme-storage/appdata/Hermes-Agent` und verwendet Athenas
-  Router als Modell-Backend.
+- Oberfläche und Agent: Hermes auf Unraid unter
+  `/mnt/nvme-storage/appdata/Hermes-Agent`. Athena stellt dafür nur die
+  OpenAI-kompatible Router-API bereit. OpenWebUI ist abgeschaltetes Rückfallnetz.
 - Inferenz: genau ein aktives llama.cpp-Textprofil; der Router wechselt bei
   Bedarf zwischen Fast, Medium, Large, Ultra und Uncensored.
 
@@ -35,16 +35,11 @@ Migration höchstens als gekennzeichnetes Archiv existieren.
 
 ## Container-Prinzip
 
-Ein eigener Container ist sinnvoll, wenn ein Dienst eigene Abhängigkeiten,
-Zugangsdaten oder eine eigene Fehlergrenze hat. Deshalb bleiben die fachlichen
-MCPs getrennt, beispielsweise Home Assistant, Unraid, ARR, Navidrome, Deemix,
-Web und SSH. Es gibt jedoch keine zusätzlichen MCPs für einzelne
-Zwischenschritte einer Installation.
-
-Hermes auf Unraid ist die primäre Oberfläche für längere administrative und
-agentische Aufgaben. OpenWebUI auf Athena erhält weiterhin die Werkzeuge, die
-für kurze Abfragen sinnvoll sind. Ein neuer MCP muss nicht automatisch in jede
-Oberfläche eingebunden werden; das richtet sich nach dem Auftrag.
+Athena betreibt nur Inferenz, Router, Sprache/Bild, Backup und den
+hostgebundenen Athena-Operator. Portable Fach-MCPs laufen als getrennte
+Unterprozesse im **einen MCPHub-Container auf Unraid**. Sie bleiben unter
+`/mcp/NAME` einzeln sichtbar und abschaltbar, brauchen aber nicht je einen
+Docker-Container. `config/mcp-registry.json` ist die einzige Serverliste.
 
 ## Standardablauf für Änderungen
 
@@ -63,18 +58,14 @@ Werkzeugaufruf wird höchstens einmal wiederholt.
 
 ## Neuer MCP
 
-Für einen neuen MCP sind gewöhnlich nur diese Teile nötig:
+Portable MCPs werden nach dem Skill `mcphub-deployer` in das reproduzierbare
+MCPHub-Image eingebaut. Dazu gehören gepinnte Quelle, genau ein Registry-Eintrag,
+Secret-Datei nur im Appdata, Build, Handshake und eine read-only-Probe. Danach
+wird ausschließlich MCPHub über Unraid DockerMan neu erstellt. Router, Qwen,
+WireGuard und andere Container werden nicht neu gestartet.
 
-1. Servercode und Dockerfile unter `platform/mcp/`.
-2. Ein Service im eingebundenen `platform/mcp/compose.yaml`.
-3. Eine Env-Beispieldatei unter `config/`; echte Werte nach `/etc/mike-ai`.
-4. Genau ein Eintrag in `config/mcp-registry.json`; daraus werden Hermes und
-   OpenWebUI automatisch erzeugt.
-5. Ein kleiner Test sowie ein kurzer Eintrag in dieser Datei oder in der
-   Komponentenübersicht, falls wirklich zusätzliche Erklärung nötig ist.
-
-Dann wird nur der neue MCP gebaut und gestartet. Router, Qwen, WireGuard,
-Hermes und der komplette Stack werden nicht pauschal neu gestartet.
+Nur ein Werkzeug, das Athenas Host selbst verwalten muss, gehört in den
+Athena-Operator. Es wird kein zweiter allgemeiner Terminal- oder Doku-MCP gebaut.
 
 ## Temporär oder dauerhaft
 
diff --git a/README.md b/README.md
index 50497e0..297c4de 100644
--- a/README.md
+++ b/README.md
@@ -1,9 +1,8 @@
 # Athena AI
 
-Ein reproduzierbarer Docker-Stack für Athenas lokale KI. Athena stellt Router,
-llama.cpp-Profile, OpenWebUI, Sprache und Bildgenerierung bereit. Der offizielle
-Hermes Agent und die portablen Werkzeuge laufen auf Unraid und werden dort
-zusammen mit dem übrigen Appdata gesichert.
+Ein reproduzierbarer Docker-Stack für Athenas lokale Inferenz. Athena stellt
+Router, llama.cpp-Profile, Sprache und Bildgenerierung bereit. Der offizielle
+Hermes Agent und MCPHub laufen auf Unraid und werden dort mit Appdata gesichert.
 
 ## Aufbau
 
@@ -21,11 +20,10 @@ zusammen mit dem übrigen Appdata gesichert.
 - Home Assistant, ARR, Unraid, Navidrome, Deemix und GitHub laufen gemeinsam im
   MCPHub-Container auf Unraid, bleiben aber als getrennte MCP-Server unter
   `/mcp/NAME` sichtbar, abschaltbar und unabhängig für Clients freigebbar.
-- SearXNG/TinySearch dürfen als eigene Such-Backends laufen. Der eigentliche
-  Web-MCP-Adapter zieht ebenfalls in den MCPHub, sobald der Backend-Pfad aus dem
-  Unraid-Netz verfügbar ist.
-- `config/mcp-registry.json` ist die einzige Liste der MCPs für Hermes und
-  OpenWebUI. `platform/mcp/sync-clients.py` erzeugt beide Registrierungen.
+- Hermes verwendet für allgemeine Recherche seine integrierten Webwerkzeuge;
+  der frühere Athena-Webadapter wird nicht mehr gestartet.
+- `config/mcp-registry.json` ist die einzige Liste der MCPHub-Server und ihrer
+  Client-Registrierungen. `platform/mcp/sync-clients.py` erzeugt Hermes daraus.
 - Modelle und Athena-Backups liegen auf `/data`. Hermes liegt vollständig unter
   `/mnt/nvme-storage/appdata/Hermes-Agent`; MCPHub-Zustand, Client-Schlüssel und
   MCP-Zugänge liegen unter `/mnt/nvme-storage/appdata/MCPHub`.
@@ -43,7 +41,7 @@ sudo ./install.sh --config config/install.env
 ```
 
 Das Skript installiert Docker und NVIDIA-Unterstützung, lädt die konfigurierten
-Modelle, baut den Stack und startet die benötigten Profile und MCPs.
+Modelle und startet ausschließlich Athenas Inferenz-Kern plus Operator.
 
 ## Bedienung
 
@@ -51,14 +49,16 @@ Modelle, baut den Stack und startet die benötigten Profile und MCPs.
 # Gesamten Stack anzeigen
 docker compose --env-file /etc/mike-ai/stack.env ps
 
-# Eine gezielte Athena-Komponente ausrollen
-docker compose --env-file /etc/mike-ai/stack.env up -d --build router
+# Erst anzeigen, dann eine gezielte Komponente ohne Nebenwirkungen ausrollen
+./manage.sh --dry-run deploy router
+./manage.sh deploy router
 
 # Sofortiges Datenbackup zusätzlich zum Fünf-Stunden-Zeitplan
 docker exec mike-ai-backup backup
-```
 
-OpenWebUI: `http://<WireGuard-IP>:8080`
+# Kurzer read-only Ende-zu-Ende-Test nach jedem Release
+sudo ./smoke-test.sh
+```
 
 Router-API: `http://<WireGuard-IP>:8081/v1`
 
@@ -81,6 +81,7 @@ Details, Prüfschritte und der exakte Sicherungsumfang stehen in
 
 - [`ATHENA.md`](ATHENA.md) – kurze Maschinen- und Operatoranleitung
 - [`docs/STANDARD_PROFILE_MATRIX.md`](docs/STANDARD_PROFILE_MATRIX.md) – Profile und Messwerte
+- [`docs/MCP_SERVERS.md`](docs/MCP_SERVERS.md) – automatisch erzeugte MCP-Liste
 - [`docs/RECOVERY.md`](docs/RECOVERY.md) – Backup und Neuaufbau
 
 Die MCPHub-Installation, Endpunkte und der schrittweise Rückbau der alten
diff --git a/compose.yaml b/compose.yaml
index ab124b2..e49d1ea 100644
--- a/compose.yaml
+++ b/compose.yaml
@@ -928,11 +928,9 @@ services:
       - /data/docker-backups:/archive
       - /etc/mike-ai:/backup/etc-mike-ai:ro
       - /opt/mike-ai/stack:/backup/stack:ro
-      - open-webui-data:/backup/volumes/open-webui-data:ro
       - piper-data:/backup/volumes/piper-data:ro
       - router-state:/backup/volumes/router-state:ro
       - router-images:/backup/volumes/router-images:ro
-      - tinysearch-models:/backup/volumes/tinysearch-models:ro
     security_opt: ["no-new-privileges:true"]
 
 networks:
diff --git a/config/mcp-registry.json b/config/mcp-registry.json
index 5a2daca..e864ab9 100644
--- a/config/mcp-registry.json
+++ b/config/mcp-registry.json
@@ -11,16 +11,13 @@
       "env_file": "/etc/mike-ai/mcphub-client.env",
       "key_env": "MCPHUB_BEARER_TOKEN",
       "auth_type": "bearer",
-      "timeout": 900
-    },
-    {
-      "id": "web-general-local",
-      "hermes_id": "web-general",
-      "name": "Allgemeines Web (TinySearch)",
-      "description": "Breite Websuche und Seitenabruf für beliebige öffentliche Websites. Kurz und gezielt suchen; keine vollständigen Websites rekursiv einlesen.",
-      "url": "http://tinysearch:8000/mcp",
-      "clients": ["hermes", "openwebui"],
-      "timeout": 180
+      "timeout": 900,
+      "hub": {
+        "type": "streamable-http",
+        "url": "http://192.168.1.212:8202/mcp",
+        "owner": "admin",
+        "enabled": true
+      }
     },
     {
       "id": "github-local",
@@ -33,7 +30,13 @@
       "key_env": "MCPHUB_BEARER_TOKEN",
       "auth_type": "bearer",
       "timeout": 300,
-      "functions": "github-search_repositories,github-get_file_contents,github-search_code"
+      "functions": "github-search_repositories,github-get_file_contents,github-search_code",
+      "hub": {
+        "type": "stdio",
+        "command": "/usr/local/bin/run-with-env",
+        "args": ["/run/secrets/mcphub/github.env", "--", "/usr/local/bin/github-mcp-server", "stdio", "--read-only", "--tools", "search_repositories,get_file_contents,search_code"],
+        "enabled": true
+      }
     },
     {
       "id": "homeassistant-local",
@@ -45,7 +48,15 @@
       "env_file": "/etc/mike-ai/mcphub-client.env",
       "key_env": "MCPHUB_BEARER_TOKEN",
       "auth_type": "bearer",
-      "timeout": 300
+      "timeout": 300,
+      "hub": {
+        "type": "streamable-http",
+        "secret_file": "homeassistant.env",
+        "url": "${HASS_URL}/api/hass_mcp",
+        "headers": {"Authorization": "Bearer ${HASS_TOKEN}"},
+        "owner": "admin",
+        "enabled": true
+      }
     },
     {
       "id": "arr-local",
@@ -57,7 +68,13 @@
       "env_file": "/etc/mike-ai/mcphub-client.env",
       "key_env": "MCPHUB_BEARER_TOKEN",
       "auth_type": "bearer",
-      "timeout": 600
+      "timeout": 600,
+      "hub": {
+        "type": "stdio",
+        "command": "/usr/local/bin/run-with-env",
+        "args": ["/run/secrets/mcphub/arr.env", "--", "arr-mcp", "--transport", "stdio", "--auth-type", "none"],
+        "enabled": true
+      }
     },
     {
       "id": "navidrome-local",
@@ -69,7 +86,14 @@
       "env_file": "/etc/mike-ai/mcphub-client.env",
       "key_env": "MCPHUB_BEARER_TOKEN",
       "auth_type": "bearer",
-      "timeout": 300
+      "timeout": 300,
+      "hub": {
+        "type": "stdio",
+        "command": "/usr/local/bin/run-with-env",
+        "args": ["/run/secrets/mcphub/navidrome.env", "--", "node", "/opt/casaderoll/navidrome/dist/index.js"],
+        "env": {"MCP_TRANSPORT": "stdio", "MCP_HTTP_EXPOSE": "false", "WEBUI_ENABLED": "false"},
+        "enabled": true
+      }
     },
     {
       "id": "deemix-local",
@@ -81,7 +105,14 @@
       "env_file": "/etc/mike-ai/mcphub-client.env",
       "key_env": "MCPHUB_BEARER_TOKEN",
       "auth_type": "bearer",
-      "timeout": 300
+      "timeout": 300,
+      "hub": {
+        "type": "stdio",
+        "command": "/usr/local/bin/run-with-env",
+        "args": ["/run/secrets/mcphub/deemix.env", "--", "python3", "/opt/casaderoll/mcps/deemix_mcp.py"],
+        "env": {"MCP_TRANSPORT": "stdio"},
+        "enabled": true
+      }
     },
     {
       "id": "mua",
@@ -93,7 +124,39 @@
       "env_file": "/etc/mike-ai/mcphub-client.env",
       "auth_type": "bearer",
       "clients": ["hermes", "openwebui"],
-      "timeout": 900
+      "timeout": 900,
+      "hub": {
+        "type": "streamable-http",
+        "secret_file": "mua.env",
+        "url": "${MUA_MCP_URL}",
+        "headers": {"Authorization": "Bearer ${MUA_MCP_BEARER_TOKEN}"},
+        "owner": "admin",
+        "enabled": true
+      }
+    },
+    {
+      "id": "fritzbox-local",
+      "hermes_id": "fritzbox",
+      "name": "FRITZ!Box",
+      "description": "FRITZ!Box-Status, WAN/Glasfaser, Netzwerkgeräte, WLAN und Telefonie. Zuerst die aktive WAN-Verbindung ermitteln; Änderungen nur auf ausdrücklichen Auftrag.",
+      "url": "http://192.168.1.2:8787/mcp/fritzbox",
+      "key_env": "MCPHUB_BEARER_TOKEN",
+      "env_file": "/etc/mike-ai/mcphub-client.env",
+      "auth_type": "bearer",
+      "clients": ["hermes", "openwebui"],
+      "timeout": 300,
+      "tool_include": [
+        "fritzbox-list_services",
+        "fritzbox-list_actions",
+        "fritzbox-describe_action",
+        "fritzbox-call_action"
+      ],
+      "hub": {
+        "type": "stdio",
+        "command": "/usr/local/bin/run-with-env",
+        "args": ["/run/secrets/mcphub/fritzbox.env", "--", "/opt/casaderoll/fritz-mcp"],
+        "enabled": true
+      }
     },
     {
       "id": "mua-readonly-local",
diff --git a/config/profile-matrix.json b/config/profile-matrix.json
new file mode 100644
index 0000000..3d83a5b
--- /dev/null
+++ b/config/profile-matrix.json
@@ -0,0 +1,62 @@
+{
+  "version": 1,
+  "default_profile": "medium",
+  "max_output_tokens": 8192,
+  "profiles": [
+    {
+      "id": "fast",
+      "alias": "qwen-fast",
+      "context": 76800,
+      "model_env": "FAST_MODEL_FILE",
+      "model_family": "Qwen3.8-27B IQ4 Mix",
+      "gpu_split": "5080 only",
+      "vision": true,
+      "mtp": 2,
+      "description": "Schnelles Profil für kurze Chats und zügige Werkzeugaufgaben."
+    },
+    {
+      "id": "medium",
+      "alias": "qwen-medium",
+      "context": 160000,
+      "model_env": "MEDIUM_MODEL_FILE",
+      "model_family": "Qwen3.8-27B IQ4 XS Pure",
+      "gpu_split": "90:10",
+      "vision": true,
+      "mtp": 3,
+      "description": "Ausgewogenes Standardprofil für Alltag und lange agentische Aufgaben."
+    },
+    {
+      "id": "large",
+      "alias": "qwen-large",
+      "context": 192000,
+      "model_env": "LARGE_MODEL_FILE",
+      "model_family": "Qwen3.8-27B IQ4 XS Pure",
+      "gpu_split": "86:14",
+      "vision": true,
+      "mtp": 3,
+      "description": "Großes Profil für umfangreiche Dokumente und lange technische Arbeiten."
+    },
+    {
+      "id": "ultra",
+      "alias": "qwen-ultra",
+      "context": 262144,
+      "model_env": "ULTRA_MODEL_FILE",
+      "model_family": "Qwen3.8-27B IQ4 XS Pure",
+      "gpu_split": "80:20",
+      "vision": false,
+      "mtp": 2,
+      "description": "Maximaler Textkontext; bewusst ohne Vision-Projektor."
+    },
+    {
+      "id": "uncensored",
+      "alias": "qwen-uncensored",
+      "context": 80000,
+      "model_env": "UNCENSORED_MODEL_FILE",
+      "model_family": "Qwen3.8-27B Abliterated Q4_K_M",
+      "gpu_split": "90:10",
+      "vision": true,
+      "mtp": 2,
+      "description": "Weniger restriktives Spezialprofil; Werkzeugrechte bleiben unverändert."
+    }
+  ]
+}
diff --git a/config/unraid-templates/my-MCPHub.xml b/config/unraid-templates/my-MCPHub.xml
index 802427c..e5032ec 100644
--- a/config/unraid-templates/my-MCPHub.xml
+++ b/config/unraid-templates/my-MCPHub.xml
@@ -1,7 +1,7 @@
 <?xml version="1.0"?>
 <Container version="2">
   <Name>MCPHub</Name>
-  <Repository>casaderoll/mcphub:1.0.0</Repository>
+  <Repository>casaderoll/mcphub:1.1.1</Repository>
   <Registry>https://hub.docker.com/r/samanhappy/mcphub</Registry>
   <Network>bridge</Network>
   <MyIP/>
@@ -11,7 +11,7 @@
   <Support>https://github.com/samanhappy/mcphub/issues</Support>
   <Project>https://github.com/samanhappy/mcphub</Project>
   <ReadMe>https://github.com/samanhappy/mcphub#readme</ReadMe>
-  <Overview>Zentrale MCP-Verwaltung mit Weboberfläche. Das lokale CasaDeRoll-Image basiert reproduzierbar auf MCPHub 1.0.32 und enthält die versionierten ARR-, Deemix-, Navidrome-, GitHub- und Web-MCP-Laufzeiten. Home Assistant und MUA werden als vorhandene HTTP-MCPs eingebunden. Einzelne Server bleiben unter /mcp/NAME getrennt sichtbar und schaltbar.&#13;
+  <Overview>Zentrale MCP-Verwaltung mit Weboberfläche. Das lokale CasaDeRoll-Image basiert reproduzierbar auf MCPHub 1.0.32 und enthält die versionierten ARR-, Deemix-, Navidrome-, GitHub- und FritzBox-Laufzeiten. Home Assistant, MUA und der Athena Operator werden als vorhandene HTTP-MCPs eingebunden. Einzelne Server bleiben unter /mcp/NAME getrennt sichtbar und schaltbar. Hermes nutzt für allgemeine Webrecherche seine eingebauten Werkzeuge.&#13;
 &#13;
 Weboberfläche: http://[IP]:[PORT:3000]/&#13;
 Benutzer beim ersten Start: admin&#13;
diff --git a/dev/mock_upstream.py b/dev/mock_upstream.py
index 2139e6f..1158d7a 100755
--- a/dev/mock_upstream.py
+++ b/dev/mock_upstream.py
@@ -99,6 +99,9 @@ class Handler(BaseHTTPRequestHandler):
         self.send_header("Connection", "close")
         self.end_headers()
         model = body.get("model")
+        delay = body.get("mock_stream_delay", 0.05)
+        if not isinstance(delay, (int, float)) or delay < 0 or delay > 10:
+            delay = 0.05
         for tok in ["Mock-", "Streaming", "-Antwort", f" ({model})"]:
             chunk = {
                 "id": "chatcmpl-mock",
@@ -109,9 +112,12 @@ class Handler(BaseHTTPRequestHandler):
             }
             self.wfile.write(f"data: {json.dumps(chunk)}\n\n".encode())
             self.wfile.flush()
-            time.sleep(0.05)
-        self.wfile.write(b"data: [DONE]\n\n")
-        self.wfile.flush()
+            time.sleep(delay)
+        try:
+            self.wfile.write(b"data: [DONE]\n\n")
+            self.wfile.flush()
+        except (BrokenPipeError, ConnectionResetError):
+            pass
 
     def _json(self, code: int, payload: dict) -> None:
         body = json.dumps(payload).encode()
diff --git a/dev/test_local.sh b/dev/test_local.sh
index 77ff8dd..a6d0789 100755
--- a/dev/test_local.sh
+++ b/dev/test_local.sh
@@ -183,6 +183,34 @@ echo "$RESP"
 echo "$RESP" | grep -q "data: " && echo "$RESP" | grep -q "\[DONE\]" \
   && ok "SSE-Stream mit [DONE] erhalten" || bad "Streaming"
 
+echo "== Test 4b: Client-Abbruch gibt den Router-Slot frei"
+command curl -H "Authorization: Bearer $TEST_ROUTER_KEY" -sfN \
+  "$BASE/v1/chat/completions" -H "Content-Type: application/json" \
+  -d '{"model":"qwen-fast","stream":true,"mock_stream_delay":2,"messages":[{"role":"user","content":"Abbruch"}]}' \
+  >/tmp/aborted-stream.txt 2>/dev/null &
+ABORT_PID=$!
+sleep 0.4
+kill "$ABORT_PID" 2>/dev/null || true
+wait "$ABORT_PID" 2>/dev/null || true
+for _ in $(seq 1 30); do
+  ACTIVE=$(curl -sf "$BASE/status" | python3 -c 'import json,sys; print(json.load(sys.stdin)["qwen"]["active_chats"])')
+  [[ $ACTIVE == 0 ]] && break
+  sleep 0.1
+done
+CODE=$(curl -s -o /tmp/after-abort.json -w "%{http_code}" "$BASE/v1/chat/completions" \
+  -H "Content-Type: application/json" \
+  -d '{"model":"qwen-fast","messages":[{"role":"user","content":"noch frei?"}]}')
+[[ ${ACTIVE:-1} == 0 && $CODE == 200 ]] \
+  && ok "abgebrochener Stream gibt Lease frei; Folgerequest erfolgreich" \
+  || bad "Stream-Abbruch hinterließ active_chats=${ACTIVE:-?}, HTTP $CODE"
+
+if [[ ${ROUTER_TEST_QUICK:-0} == 1 ]]; then
+  echo
+  echo "== Schnellergebnis: $PASS bestanden, $FAIL fehlgeschlagen =="
+  [[ $FAIL -eq 0 ]]
+  exit
+fi
+
 # --- 5. Tool Calls -----------------------------------------------------------------
 echo "== Test 5: Tool Calls"
 RESP=$(curl -sf "$BASE/v1/chat/completions" -H "Content-Type: application/json" \
diff --git a/dev/test_mcp_registry.py b/dev/test_mcp_registry.py
index 6e3b3d4..2158753 100644
--- a/dev/test_mcp_registry.py
+++ b/dev/test_mcp_registry.py
@@ -66,6 +66,41 @@ class RegistryTests(unittest.TestCase):
         con.close()
         self.assertEqual(ids, ["unmanaged", "one"])
 
+    def test_production_registry_has_unique_ids_and_fritzbox(self):
+        document = json.loads((ROOT / "config/mcp-registry.json").read_text())
+        ids = [item["id"] for item in document["servers"]]
+        hermes_ids = [
+            item.get("hermes_id", item["id"])
+            for item in document["servers"] if "hermes" in item.get("clients", [])
+        ]
+        self.assertEqual(len(ids), len(set(ids)))
+        self.assertEqual(len(hermes_ids), len(set(hermes_ids)))
+        fritz = next(item for item in document["servers"] if item["id"] == "fritzbox-local")
+        self.assertEqual(fritz["hub"]["type"], "stdio")
+        self.assertIn("fritz-mcp", fritz["hub"]["args"][-1])
+        self.assertEqual(len(fritz["tool_include"]), 4)
+
+    def test_hermes_tool_filter_is_generated(self):
+        item = {
+            "id": "wide", "name": "Wide", "description": "Test",
+            "url": "http://wide/mcp", "clients": ["hermes"],
+            "tool_include": ["list", "describe", "call"],
+        }
+        block = self.module.hermes_block([item])
+        self.assertIn("    tools:\n      include:", block)
+        self.assertIn('        - "describe"', block)
+
+    def test_raw_mcphub_token_can_replace_host_specific_env_file(self):
+        item = {
+            "id": "hub", "name": "Hub", "description": "Test",
+            "url": "http://hub/mcp/test", "clients": ["hermes"],
+            "env_file": "/missing/client.env",
+            "key_env": "MCPHUB_BEARER_TOKEN",
+        }
+        self.module.CLIENT_TOKEN = "local-token"
+        self.assertTrue(self.module.enabled(item))
+        self.assertEqual(self.module.resolved(item), ("http://hub/mcp/test", "local-token"))
+
 
 if __name__ == "__main__":
     unittest.main()
diff --git a/dev/test_mcphub_settings.py b/dev/test_mcphub_settings.py
new file mode 100644
index 0000000..fd99e6e
--- /dev/null
+++ b/dev/test_mcphub_settings.py
@@ -0,0 +1,81 @@
+#!/usr/bin/env python3
+"""Regression tests for declarative, update-safe MCPHub settings."""
+
+from __future__ import annotations
+
+import importlib.util
+import json
+import os
+import tempfile
+import unittest
+from pathlib import Path
+
+
+ROOT = Path(__file__).parents[1]
+SOURCE = ROOT / "platform/mcphub/configure-settings.py"
+
+
+def load_module():
+    spec = importlib.util.spec_from_file_location("configure_settings", SOURCE)
+    module = importlib.util.module_from_spec(spec)
+    assert spec.loader
+    spec.loader.exec_module(module)
+    return module
+
+
+class MCPHubSettingsTests(unittest.TestCase):
+    def setUp(self):
+        self.module = load_module()
+        self.temp = tempfile.TemporaryDirectory()
+        self.root = Path(self.temp.name)
+        self.secrets = self.root / "secrets"
+        self.secrets.mkdir()
+        (self.secrets / "remote.env").write_text(
+            "REMOTE_URL=http://example.test/mcp\nTOKEN=secret-value\n",
+            encoding="utf-8",
+        )
+        self.registry = self.root / "registry.json"
+        self.registry.write_text(json.dumps({
+            "version": 1,
+            "servers": [
+                {
+                    "id": "remote-local",
+                    "hermes_id": "remote",
+                    "hub": {
+                        "type": "streamable-http",
+                        "secret_file": "remote.env",
+                        "url": "${REMOTE_URL}",
+                        "headers": {"Authorization": "Bearer ${TOKEN}"},
+                        "enabled": True,
+                    },
+                },
+                {"id": "client-only", "url": "http://unused/mcp"},
+            ],
+        }), encoding="utf-8")
+
+    def tearDown(self):
+        self.temp.cleanup()
+
+    def test_registry_renders_only_hub_servers_and_expands_secrets(self):
+        servers = self.module.registry_servers(self.registry, self.secrets, {})
+        self.assertEqual(list(servers), ["remote"])
+        self.assertEqual(servers["remote"]["url"], "http://example.test/mcp")
+        self.assertEqual(
+            servers["remote"]["headers"]["Authorization"],
+            "Bearer secret-value",
+        )
+
+    def test_existing_enabled_toggle_survives_reconciliation(self):
+        servers = self.module.registry_servers(
+            self.registry, self.secrets, {"remote": {"enabled": False}}
+        )
+        self.assertFalse(servers["remote"]["enabled"])
+
+    def test_missing_secret_fails_closed(self):
+        os.unlink(self.secrets / "remote.env")
+        with self.assertRaises(SystemExit):
+            self.module.registry_servers(self.registry, self.secrets, {})
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/docs/MCP_SERVERS.md b/docs/MCP_SERVERS.md
new file mode 100644
index 0000000..07fb947
--- /dev/null
+++ b/docs/MCP_SERVERS.md
@@ -0,0 +1,16 @@
+# MCP-Server
+
+Diese Datei wird aus `config/mcp-registry.json` erzeugt. Änderungen gehören nur in die JSON-Registry.
+
+| Server | Hermes-ID | Endpunkt | Clients | Werkzeuge |
+|---|---|---|---|---:|
+| Athena Operator | `athena-operator` | `http://192.168.1.2:8787/mcp/athena-operator` | hermes, openwebui | alle |
+| GitHub (offiziell, read-only) | `github` | `http://192.168.1.2:8787/mcp/github` | hermes, openwebui | alle |
+| Home Assistant | `homeassistant-admin` | `http://192.168.1.2:8787/mcp/homeassistant` | hermes, openwebui | alle |
+| Sonarr und Radarr | `arr` | `http://192.168.1.2:8787/mcp/arr` | hermes, openwebui | alle |
+| Navidrome | `navidrome` | `http://192.168.1.2:8787/mcp/navidrome` | hermes, openwebui | alle |
+| Deemix | `deemix` | `http://192.168.1.2:8787/mcp/deemix` | hermes, openwebui | alle |
+| MUA (Unraid-Verwaltung) | `unraid` | `http://192.168.1.2:8787/mcp/unraid` | hermes, openwebui | alle |
+| FRITZ!Box | `fritzbox` | `http://192.168.1.2:8787/mcp/fritzbox` | hermes, openwebui | 4 |
+
+Allgemeine Webrecherche ist ein eingebautes Hermes-Werkzeug und kein MCPHub-Server.
diff --git a/docs/RECOVERY.md b/docs/RECOVERY.md
index e14cfa9..f210d2c 100644
--- a/docs/RECOVERY.md
+++ b/docs/RECOVERY.md
@@ -3,18 +3,14 @@
 ## Was automatisch gesichert wird
 
 Der Container `mike-ai-backup` erstellt alle fünf Stunden ein komprimiertes
-Archiv unter `/data/docker-backups` und behält 14 Tage. Während der kurzen
-Sicherung wird nur OpenWebUI angehalten, damit seine SQLite-Datenbank konsistent
-ist. Netzwerk, WireGuard, Router und Qwen bleiben erreichbar. Hermes läuft
-unabhängig davon auf Unraid weiter.
+Archiv unter `/data/docker-backups` und behält 14 Tage. Netzwerk, WireGuard,
+Router und Qwen bleiben dabei erreichbar. Hermes läuft unabhängig auf Unraid.
 
 Enthalten sind:
 
 - `/etc/mike-ai` mit lokalen Konfigurationen und Secrets
-- OpenWebUI-Daten
 - Router-Zustand und Router-Bildablage
 - Piper-Daten
-- TinySearch-Modellcache
 - ein Quellbaum-Snapshot als zusätzliche Bequemlichkeit
 
 Nicht kopiert werden `/data/models`, die nur noch als Rückfall vorhandene Kopie
@@ -84,7 +80,7 @@ docker compose --env-file /etc/mike-ai/stack.env ps
 test -s /data/docker-backups/athena-latest.tar.gz
 ```
 
-Danach einen OpenWebUI-Login, einen Router-Request und je einen read-only
+Danach einen Router-Request, einen Hermes-Zweiturn-Chat und je einen read-only
 MCP-Aufruf über `http://UNRAID-IP:8787/mcp/NAME` testen. Alte Recovery-Koffer
 sind für Neuinstallationen nicht mehr erforderlich; Git, Athenas Datenbackup
 und das Unraid-Appdata-Backup bilden die Wiederherstellung.
diff --git a/docs/STANDARD_PROFILE_MATRIX.md b/docs/STANDARD_PROFILE_MATRIX.md
index 6aa7ffc..e2932fb 100644
--- a/docs/STANDARD_PROFILE_MATRIX.md
+++ b/docs/STANDARD_PROFILE_MATRIX.md
@@ -1,50 +1,21 @@
-# Verbindliche Standard-Profilmatrix
+# Profilmatrix
 
-Stand: 22. August 2026. Diese fünf Profile sind die Produktionsmatrix für
-Athena. Änderungen gelten erst nach einem vergleichbaren synthetischen Test
-und einer bewussten Aktualisierung dieser Datei.
+Diese Datei wird aus `config/profile-matrix.json` erzeugt. Änderungen gehören nur in die JSON-Matrix.
 
-| Profil | Virtuelles Modell | GGUF | Kontext | GPUs / Split | MTP | Vision | gemessene kurze Ausgabe |
-|---|---|---|---:|---|---:|---|---:|
-| Fast | `qwen-fast` | IQ4-MIX | 76.800 | RTX 5080 | 2 | ja, Projektor auf RTX 3060 | 85,5 Tok/s |
-| **Medium (Default)** | `qwen-medium` | IQ4_XS Pure | 160.000 | RTX 5080 + RTX 3060, 90:10 | 3 | ja, Projektor auf RTX 3060 | 77,2 Tok/s |
-| Large | `qwen-large` | IQ4_XS Pure | 192.000 | RTX 5080 + RTX 3060, 86:14 | 3 | ja, Projektor auf RTX 3060 | 75,3 Tok/s |
-| Ultra | `qwen-ultra` | IQ4_XS Pure | 262.144 | RTX 5080 + RTX 3060, 80:20 | 2 | nein, text-only | 68,2 Tok/s |
-| Uncensored | `qwen-uncensored` | Abliterated Q4_K_M | 80.000 | RTX 5080 + RTX 3060, 90:10 | 2 | ja, eigener Projektor auf RTX 3060 | 52,2 Tok/s |
+Standardprofil: **medium** · globales Ausgabelimit: **8192 Token**
 
-## Standardverhalten
+| Profil | API-Alias | Kontext | Modell | GPU-Verteilung | Vision | MTP |
+|---|---|---:|---|---|---|---:|
+| fast | `qwen-fast` | 76,800 | Qwen3.8-27B IQ4 Mix | 5080 only | ja | 2 |
+| medium | `qwen-medium` | 160,000 | Qwen3.8-27B IQ4 XS Pure | 90:10 | ja | 3 |
+| large | `qwen-large` | 192,000 | Qwen3.8-27B IQ4 XS Pure | 86:14 | ja | 3 |
+| ultra | `qwen-ultra` | 262,144 | Qwen3.8-27B IQ4 XS Pure | 80:20 | nein | 2 |
+| uncensored | `qwen-uncensored` | 80,000 | Qwen3.8-27B Abliterated Q4_K_M | 90:10 | ja | 2 |
 
-- Medium ist nach Neuinstallation und bewusstem Plattformstart das aktive
-  Standardprofil.
-- Open WebUI erhält `mikeai-medium` als Standardmodell.
-- Im OpenWebUI-Arbeitsbereich ist dieses auf `qwen-medium` basierende Preset
-  `mikeai-medium` die sichtbare Standardauswahl; die fünf rohen `qwen-*`-Aliase
-  bleiben ausgeblendet.
-- Ein explizit gewähltes anderes Modell löst den zugehörigen Containerwechsel
-  aus; es ist immer nur ein Inferenzcontainer aktiv.
-- Ultra reserviert den verfügbaren Speicher für nativen 256K-Textkontext und
-  lädt deshalb keinen Vision-Projektor.
-- Uncensored ist ein bewusst gewähltes Spezialprofil mit reduzierter
-  Verweigerungsneigung. Es ändert weder Router-Authentifizierung noch
-  Tool-Rechte, Bestätigungsregeln oder Secret-Filter und ist nicht Default.
-- Das gesonderte Experimentalprofil gehört nicht zur Benutzer-Matrix und wird
-  in Open WebUI nicht als reguläres Modell angeboten.
+## Zweck
 
-## Hermes-Kontextpflege
-
-Hermes verwendet das jeweils ausgewählte 27B-Profil auch für die
-Kontextzusammenfassung. Micro-Compaction übernimmt alle fünf abgeschlossenen
-Turns einen älteren Assistenten-/Werkzeugabschnitt in die laufende
-Zusammenfassung. Die normale schwellenbasierte Kompression bleibt als
-Rückfallebene aktiv. Ein separates kleines Kompressionsmodell wird nicht
-installiert, weil 2B/4B-Tests exakte technische Zustände verlieren konnten.
-
-## Nachweise
-
-- Fast 76,8K: synthetischer Referenzlauf, 85,50 Tok/s.
-- Medium 160K: Pure 90:10 mit MTP3, 77,22 Tok/s.
-- Large 192K: Pure 86:14 mit MTP3, 75,28 Tok/s.
-- Ultra 256K: Pure 80:20 mit MTP2, 68,19 Tok/s; 220.190 Tokens
-  erfolgreich verarbeitet und Sentinel korrekt wiedergefunden.
-- Uncensored 80K: Abliterated Q4_K_M 90:10 mit MTP2, 52,2 Tok/s; 60K-Fülltest,
-  Vision und synthetischer Logik-/Sicherheitslauf bestanden.
+- **fast**: Schnelles Profil für kurze Chats und zügige Werkzeugaufgaben.
+- **medium**: Ausgewogenes Standardprofil für Alltag und lange agentische Aufgaben.
+- **large**: Großes Profil für umfangreiche Dokumente und lange technische Arbeiten.
+- **ultra**: Maximaler Textkontext; bewusst ohne Vision-Projektor.
+- **uncensored**: Weniger restriktives Spezialprofil; Werkzeugrechte bleiben unverändert.
diff --git a/install.sh b/install.sh
index 8f81b43..2d23869 100755
--- a/install.sh
+++ b/install.sh
@@ -485,16 +485,14 @@ build_and_start() {
       "from huggingface_hub import snapshot_download; snapshot_download('black-forest-labs/FLUX.2-klein-4B', revision='303481f0390afb112393f9d77e8f0be72fcefeb7', local_dir='/download')"
     chmod -R a-w "${FLUX_MODEL_DIR:-/data/models/FLUX.2-klein-4B}"
   fi
-  # Creates the shared internal tools network before Open WebUI is created.
-  # Web search always starts; HA/ARR/Unraid only start when their root-only
-  # secret files and required local artifacts are present.
+  # Creates the tools network and deploys the only host-bound MCP: Operator.
+  # Portable MCPs and Hermes live on Unraid and are restored through Appdata.
   "$STACK_DIR/platform/mcp/install-tools.sh"
-  "$STACK_DIR/platform/hermes/install-hermes.sh"
   docker compose --env-file "$SECRETS_DIR/stack.env" --profile inference create \
     llama-fast llama-medium llama-large llama-ultra llama-experimental
   docker compose --env-file "$SECRETS_DIR/stack.env" --profile image create flux-worker
   docker compose --env-file "$SECRETS_DIR/stack.env" up -d --build \
-    xtts piper tts-gateway profile-controller router open-webui hermes
+    wireguard-gateway xtts piper tts-gateway profile-controller router backup
   if [[ ${WIREGUARD_MODE:-container} == container ]]; then
     systemctl restart mike-ai-container-vpn-guard.service
   fi
@@ -541,16 +539,6 @@ PY
   grep -Fxq 'INSTALL_READINESS_OK' <<<"$activation_output" || \
     die "Medium-Standardprofil lieferte keinen bestätigten Readiness-Marker"
 
-  log "Hermes-Profile aus der Standardmatrix anlegen"
-  "$STACK_DIR/platform/hermes/install-profiles.sh"
-
-  log "OpenWebUI-Filter und MCP-Registry synchronisieren"
-  "$STACK_DIR/platform/openwebui/install-filters.sh"
-
-  if [[ ${INSTALL_HERMES_WEBUI:-true} == true ]]; then
-    log "Entfernbare Hermes Community-WebUI installieren"
-    "$STACK_DIR/platform/hermes/install-webui.sh"
-  fi
 }
 
 hostnamectl set-hostname "$AI_HOSTNAME"
@@ -578,9 +566,8 @@ else
   vpn_address=$AI_BIND_ADDRESS
 fi
 cat <<EOF
-OpenWebUI:  http://${vpn_address}:8080
 Router-API: http://${vpn_address}:8081
-Hermes:     http://${vpn_address}:9119
+Hermes läuft mit MCPHub auf Unraid und nutzt diese Router-API.
 
 Die geheimen Schlüssel liegen ausschließlich unter $SECRETS_DIR (0600).
 Im Container-Modus stammen Peer, Adresse und Heimrouten vollständig aus dem
diff --git a/manage.sh b/manage.sh
new file mode 100755
index 0000000..d75c8a2
--- /dev/null
+++ b/manage.sh
@@ -0,0 +1,71 @@
+#!/usr/bin/env bash
+# Small, predictable operator entrypoint for Athena's Docker stack.
+set -Eeuo pipefail
+
+ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+ENV_FILE=${STACK_ENV:-/etc/mike-ai/stack.env}
+DRY_RUN=false
+
+usage() {
+  cat <<'EOF'
+Usage:
+  ./manage.sh [--dry-run] validate
+  ./manage.sh [--dry-run] deploy core
+  ./manage.sh [--dry-run] deploy SERVICE [SERVICE ...]
+  ./manage.sh [--dry-run] stop-legacy
+
+`deploy SERVICE` deliberately uses --no-deps. It never recreates unrelated
+containers. `deploy core` is the complete Athena runtime and may start missing
+dependencies, but does not start OpenWebUI, Hermes or portable Fach-MCPs.
+EOF
+}
+
+if [[ ${1:-} == --dry-run ]]; then
+  DRY_RUN=true
+  shift
+fi
+
+run() {
+  printf '+ '
+  printf '%q ' "$@"
+  printf '\n'
+  $DRY_RUN || "$@"
+}
+
+compose=(docker compose --env-file "$ENV_FILE" -f "$ROOT_DIR/compose.yaml")
+command=${1:-}
+shift || true
+
+case "$command" in
+  validate)
+    run python3 "$ROOT_DIR/platform/scripts/sync-profile-matrix.py" --check
+    run python3 "$ROOT_DIR/platform/scripts/render-mcp-registry.py" --check
+    run "${compose[@]}" config --quiet
+    ;;
+  deploy)
+    [[ $# -gt 0 ]] || { usage >&2; exit 2; }
+    run python3 "$ROOT_DIR/platform/scripts/sync-profile-matrix.py" --check
+    run python3 "$ROOT_DIR/platform/scripts/render-mcp-registry.py" --check
+    if [[ $1 == core ]]; then
+      shift
+      [[ $# -eq 0 ]] || { echo "core akzeptiert keine weiteren Services" >&2; exit 2; }
+      run "$ROOT_DIR/platform/mcp/install-tools.sh"
+      run "${compose[@]}" up -d --build \
+        wireguard-gateway piper xtts tts-gateway profile-controller router backup
+    else
+      run "${compose[@]}" up -d --build --no-deps "$@"
+    fi
+    ;;
+  stop-legacy)
+    # Reversible cleanup: stop only obsolete Athena frontends/tool backends.
+    for name in mike-ai-open-webui mike-ai-hermes mike-ai-hermes-webui \
+      mike-ai-hermes-webui-vpn-proxy mike-ai-mcp-web mike-ai-tools-searxng \
+      mike-ai-tools-tinysearch mike-ai-mcp-platform-context; do
+      if docker inspect "$name" >/dev/null 2>&1; then
+        run docker stop "$name"
+      fi
+    done
+    ;;
+  -h|--help|"") usage ;;
+  *) echo "Unbekannter Befehl: $command" >&2; usage >&2; exit 2 ;;
+esac
diff --git a/platform/checks/verify-platform.sh b/platform/checks/verify-platform.sh
index 02a75cd..5a21de1 100755
--- a/platform/checks/verify-platform.sh
+++ b/platform/checks/verify-platform.sh
@@ -1,102 +1,6 @@
 #!/usr/bin/env bash
-set -uo pipefail
+# Compatibility entrypoint: there is only one authoritative release check.
+set -Eeuo pipefail
 
-PASS=0
-WARN=0
-FAIL=0
-
-pass() { printf 'PASS  %s\n' "$*"; PASS=$((PASS + 1)); }
-warn() { printf 'WARN  %s\n' "$*"; WARN=$((WARN + 1)); }
-fail() { printf 'FAIL  %s\n' "$*"; FAIL=$((FAIL + 1)); }
-
-echo "== Local AI platform verification =="
-
-ROOT_USE="$(df -P / | awk 'NR==2 {gsub(/%/, "", $5); print $5}')"
-if [[ -n "$ROOT_USE" && "$ROOT_USE" -lt 85 ]]; then
-  pass "Systempartition bei ${ROOT_USE}%"
-else
-  fail "Systempartition bei ${ROOT_USE:-unbekannt}% (Ziel: unter 85%)"
-fi
-
-if command -v nvidia-smi >/dev/null 2>&1; then
-  GPU_LIST="$(nvidia-smi --query-gpu=name,memory.total --format=csv,noheader 2>/dev/null)"
-  GPU_COUNT="$(printf '%s\n' "$GPU_LIST" | sed '/^[[:space:]]*$/d' | wc -l)"
-  if [[ "$GPU_COUNT" -ge 2 ]]; then
-    pass "beide GPUs erkannt: $(printf '%s' "$GPU_LIST" | paste -sd ' | ' -)"
-  elif [[ "$GPU_COUNT" -eq 1 ]]; then
-    fail "nur eine von zwei erwarteten GPUs erkannt: $GPU_LIST"
-  else
-    fail "nvidia-smi liefert keine GPU"
-  fi
-else
-  fail "nvidia-smi fehlt"
-fi
-
-if [[ -e /sys/class/net/lan0 ]] && \
-   [[ "$(< /sys/class/net/lan0/address)" == "58:11:22:bb:ad:0c" ]] && \
-   [[ "$(< /sys/class/net/lan0/operstate)" == "up" ]]; then
-  pass "stabiles LAN-Interface lan0 aktiv (58:11:22:bb:ad:0c)"
-else
-  fail "stabiles LAN-Interface lan0 fehlt, ist down oder hat die falsche MAC"
-fi
-
-container_healthy() {
-  local name=$1 state health
-  state="$(docker inspect --format '{{.State.Status}}' "$name" 2>/dev/null || true)"
-  health="$(docker inspect --format '{{if .State.Health}}{{.State.Health.Status}}{{end}}' "$name" 2>/dev/null || true)"
-  [[ $state == running && ( -z $health || $health == healthy ) ]]
-}
-
-for container in mike-ai-profile-controller mike-ai-router mike-ai-xtts \
-  mike-ai-piper mike-ai-tts-gateway mike-ai-open-webui; do
-  if container_healthy "$container"; then
-    pass "$container gesund"
-  else
-    fail "$container fehlt oder ist nicht gesund"
-  fi
-done
-
-ACTIVE_LLAMA="$(docker ps --format '{{.Names}}' | grep -Ec '^mike-ai-llama-(fast|medium|large|ultra|uncensored|experimental)$' || true)"
-if [[ $ACTIVE_LLAMA -eq 1 ]]; then
-  pass "exakt ein llama.cpp-Profil aktiv"
-else
-  fail "$ACTIVE_LLAMA llama.cpp-Profile aktiv (erwartet: 1)"
-fi
-
-for optional in mike-ai-mcp-web mike-ai-mcp-homeassistant mike-ai-mcp-arr \
-  mike-ai-mcp-github mike-ai-mcp-athena-operator mike-ai-backup; do
-  if container_healthy "$optional"; then
-    pass "$optional aktiv"
-  else
-    warn "$optional nicht aktiv oder nicht installiert"
-  fi
-done
-
-if docker ps -a --format '{{.Names}}' | grep -qx 'mike-ai-mcp-unraid-official'; then
-  fail "veralteter GraphQL-basierter Unraid-MCP ist noch vorhanden"
-else
-  pass "kein GraphQL-basierter Unraid-MCP vorhanden"
-fi
-
-if docker exec mike-ai-router python -c \
-  "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8081/health', timeout=3)" \
-  >/dev/null 2>&1; then
-  pass "Router-Liveness intern erreichbar"
-else
-  fail "Router-Liveness intern nicht erreichbar"
-fi
-
-if systemctl list-unit-files --no-legend 2>/dev/null | grep -Eq '(vision-rx|whisper-rx|granite-rx).*enabled'; then
-  warn "Aktivierte RX-Altlast gefunden"
-else
-  pass "Keine aktivierte RX-Altlast"
-fi
-
-if systemctl list-unit-files --no-legend 2>/dev/null | grep -E '(benchmark|race).*enabled' >/dev/null; then
-  warn "Automatisch aktivierter Benchmark-/Race-Dienst gefunden"
-else
-  pass "Keine automatisch aktivierten Benchmarks"
-fi
-
-printf '\nErgebnis: %d PASS, %d WARN, %d FAIL\n' "$PASS" "$WARN" "$FAIL"
-[[ "$FAIL" -eq 0 ]]
+ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
+exec "$ROOT_DIR/smoke-test.sh" "$@"
diff --git a/platform/hermes/config.yaml b/platform/hermes/config.yaml
index 0dd468c..e225975 100644
--- a/platform/hermes/config.yaml
+++ b/platform/hermes/config.yaml
@@ -59,7 +59,7 @@ agent:
   # Keep enough deliberation for tool choice while avoiding the provider's
   # unbounded `auto` reasoning mode on ordinary turns. Users can still raise it
   # per session with /reasoning.
-  reasoning_effort: "minimal"
+  reasoning_effort: "low"
   gateway_timeout: 3600
   session_stall_timeout: 600
   tool_loop_guardrails:
@@ -77,7 +77,9 @@ agent:
       max_web_searches: 8
       max_subagents: 8
 
-# Keep ample room for long agent work. Compression starts at 82% of whichever
+# Keep ample room for long agent work. Compression starts late; deterministic
+# tool-result pruning and rolling micro-compaction keep it from getting there
+# during normal jobs.
 # router profile is selected. Keep the result compact enough that a local
 # model does not spend many minutes generating the handoff.
 compression:
@@ -89,19 +91,26 @@ compression:
   micro_compact_every_n_turns: 5
   micro_compact_defrag_threshold_tokens: 2000
   progress_notices: true
-  threshold: 0.82
+  threshold: 0.95
   target_ratio: 0.15
+  max_attempts: 1
   tail_mode: "lean"
   protect_last_n: 20
   protect_first_n: 0
-  proactive_prune_tokens: 50000
-  proactive_prune_min_result_chars: 4000
-  proactive_prune_min_reclaim_tokens: 4096
+  proactive_prune_tokens: 32000
+  proactive_prune_min_result_chars: 2000
+  proactive_prune_min_reclaim_tokens: 2048
+  context_timeout_seconds: 45
   # A failed local summarizer must not block an interactive client for ten
   # minutes. Continue without dropping messages after two minutes.
   context_total_ceiling_seconds: 120
 
 auxiliary:
+  compression:
+    provider: "main"
+    reasoning_effort: "none"
+    timeout: 120
+    max_concurrency: 1
   # Session names are cosmetic and used to create a second concurrent LLM
   # request after every first reply. On a single inference slot this blocks the
   # actual chat, so keep the original timestamp/session id instead.
diff --git a/platform/hermes/install-profiles.sh b/platform/hermes/install-profiles.sh
index af883e3..d24199e 100755
--- a/platform/hermes/install-profiles.sh
+++ b/platform/hermes/install-profiles.sh
@@ -1,23 +1,26 @@
 #!/usr/bin/env bash
 set -Eeuo pipefail
 
-HERMES_CONTAINER=${HERMES_CONTAINER:-mike-ai-hermes}
+HERMES_CONTAINER=${HERMES_CONTAINER:-Hermes-Agent}
 SECRETS_DIR=${SECRETS_DIR:-/etc/mike-ai}
-HERMES_DATA_DIR=${HERMES_DATA_DIR:-/data/hermes}
+HERMES_DATA_DIR=${HERMES_DATA_DIR:-/mnt/nvme-storage/appdata/Hermes-Agent}
+STACK_DIR=${STACK_DIR:-$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)}
+PROFILE_MATRIX=${PROFILE_MATRIX:-$STACK_DIR/config/profile-matrix.json}
 
 die() { printf 'FEHLER: %s\n' "$*" >&2; exit 1; }
 docker inspect "$HERMES_CONTAINER" >/dev/null 2>&1 || \
   die "Hermes-Container fehlt: $HERMES_CONTAINER"
+[[ -s $PROFILE_MATRIX ]] || die "Profilmatrix fehlt: $PROFILE_MATRIX"
 
 deadline=$((SECONDS + 180))
 until [[ $(docker inspect --format '{{if .State.Health}}{{.State.Health.Status}}{{else}}{{.State.Status}}{{end}}' \
-  "$HERMES_CONTAINER" 2>/dev/null || true) == healthy ]]; do
+  "$HERMES_CONTAINER" 2>/dev/null || true) =~ ^(healthy|running)$ ]]; do
   (( SECONDS < deadline )) || die "Hermes wurde nicht rechtzeitig gesund."
   sleep 2
 done
 
 create_profile() {
-  local name=$1 model=$2 context=$3 description=$4
+  local name=$1 model=$2 context=$3 description=$4 max_tokens=$5
   if ! docker exec "$HERMES_CONTAINER" hermes profile show "$name" >/dev/null 2>&1; then
     docker exec "$HERMES_CONTAINER" hermes profile create "$name" \
       --clone-from default --description "$description"
@@ -26,23 +29,32 @@ create_profile() {
   docker exec "$HERMES_CONTAINER" hermes -p "$name" config set model.context_length "$context"
   # These are platform-wide latency and loop safeguards, not model-specific
   # tuning. Enforce them on old profiles as well as newly cloned profiles.
-  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set model.max_tokens 8192
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set model.max_tokens "$max_tokens"
   docker exec "$HERMES_CONTAINER" hermes -p "$name" config set agent.max_turns 64
-  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set agent.reasoning_effort minimal
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set agent.reasoning_effort low
   docker exec "$HERMES_CONTAINER" hermes -p "$name" config set display.show_reasoning false
   docker exec "$HERMES_CONTAINER" hermes -p "$name" config set auxiliary.title_generation.enabled false
   docker exec "$HERMES_CONTAINER" hermes -p "$name" config unset compression.threshold_tokens || true
-  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set compression.threshold 0.82
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set compression.threshold 0.95
   docker exec "$HERMES_CONTAINER" hermes -p "$name" config set compression.target_ratio 0.15
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set compression.max_attempts 1
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set compression.context_timeout_seconds 45
   docker exec "$HERMES_CONTAINER" hermes -p "$name" config set compression.context_total_ceiling_seconds 120
   docker exec "$HERMES_CONTAINER" hermes -p "$name" config set compression.protect_last_n 20
   docker exec "$HERMES_CONTAINER" hermes -p "$name" config set compression.protect_first_n 0
   docker exec "$HERMES_CONTAINER" hermes -p "$name" config set compression.micro_compact true
   docker exec "$HERMES_CONTAINER" hermes -p "$name" config set compression.micro_compact_every_n_turns 5
   docker exec "$HERMES_CONTAINER" hermes -p "$name" config set compression.micro_compact_defrag_threshold_tokens 2000
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set compression.proactive_prune_tokens 32000
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set compression.proactive_prune_min_result_chars 2000
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set compression.proactive_prune_min_reclaim_tokens 2048
   # The selected 27B profile performs production compaction. A separate small
   # compressor was removed after it lost exact technical state in benchmarks.
   docker exec "$HERMES_CONTAINER" hermes -p "$name" config unset auxiliary.compression || true
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set auxiliary.compression.provider main
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set auxiliary.compression.reasoning_effort none
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set auxiliary.compression.timeout 120
+  docker exec "$HERMES_CONTAINER" hermes -p "$name" config set auxiliary.compression.max_concurrency 1
   docker exec "$HERMES_CONTAINER" hermes -p "$name" config set platform_toolsets.cli \
     '["web","terminal","file","skills","todo","memory","vision","tts"]'
   [[ $(docker exec "$HERMES_CONTAINER" hermes -p "$name" config get model.default) == "$model" ]] || \
@@ -51,51 +63,32 @@ create_profile() {
     die "Kontext von Profil $name konnte nicht verifiziert werden."
 }
 
-create_profile fast qwen-fast 76800 \
-  "Schnelles Qwen3.8-27B-Profil mit 76,8K Kontext fuer kurze Chats und schnelle Aufgaben."
-create_profile medium qwen-medium 160000 \
-  "Ausgewogenes Qwen3.8-27B-Standardprofil mit 160K Kontext fuer Alltag und agentische Aufgaben."
-create_profile large qwen-large 192000 \
-  "Grosses Qwen3.8-27B-Profil mit 192K Kontext fuer umfangreiche Dokumente und lange Aufgaben."
-create_profile ultra qwen-ultra 262144 \
-  "Maximales Qwen3.8-27B-Profil mit 262K Kontext fuer sehr grosse Kontexte; langsamer als die Standardprofile."
-create_profile uncensored qwen-uncensored 80000 \
-  "Unzensiertes Qwen3.8-27B-Profil mit 80K Kontext fuer spezielle Anfragen."
+while IFS=$'\t' read -r name model context description max_tokens; do
+  create_profile "$name" "$model" "$context" "$description" "$max_tokens"
+done < <(jq -r --argjson max "$(jq '.max_output_tokens' "$PROFILE_MATRIX")" \
+  '.profiles[] | [.id,.alias,(.context|tostring),(.description|gsub("[\\t\\n]";" ")),($max|tostring)] | @tsv' \
+  "$PROFILE_MATRIX")
 
-# Existing profiles may predate managed secret rendering and therefore contain
-# the literal ${ROUTER_API_KEY}. Repair only that exact placeholder; never log
-# or commit the secret itself.
-[[ -s $SECRETS_DIR/router-api-key ]] || die "Router-API-Key fehlt."
-router_key=$(<"$SECRETS_DIR/router-api-key")
-ROUTER_API_KEY="$router_key" HERMES_DATA_DIR="$HERMES_DATA_DIR" python3 <<'PY'
-import os
-import pathlib
-
-root = pathlib.Path(os.environ["HERMES_DATA_DIR"]) / "profiles"
-placeholder = "${ROUTER_API_KEY}"
-paths = [pathlib.Path(os.environ["HERMES_DATA_DIR"]) / "config.yaml"]
-paths.extend(sorted(root.glob("*/config.yaml")))
-for path in paths:
-    if not path.exists():
-        continue
-    text = path.read_text()
-    if placeholder in text:
-        text = text.replace(placeholder, os.environ["ROUTER_API_KEY"], 1)
-    # Avoid collision with Hermes' disabled built-in `homeassistant` toolset.
-    # The collision filters the healthy external MCP out of agent snapshots.
-    text = text.replace(
-        "\n  homeassistant:\n    url: http://mcp-homeassistant:8000/mcp\n",
-        "\n  homeassistant-admin:\n    url: http://mcp-homeassistant:8000/mcp\n",
-    )
-    path.write_text(text)
-PY
-
-sync_args=(--registry "${STACK_DIR:-/opt/mike-ai/stack}/config/mcp-registry.json")
+# The Unraid host deliberately has no system Python. Run the small declarative
+# client renderer in the already version-pinned MCPHub image instead of adding
+# host dependencies.
+sync_args=(--registry /stack/config/mcp-registry.json)
+mcphub_token=${MCPHUB_TOKEN_FILE:-/mnt/nvme-storage/appdata/MCPHub/client-token}
+token_mount=()
+if [[ -s $mcphub_token ]]; then
+  token_mount=(-v "$mcphub_token:/run/input/mcphub-token:ro")
+  sync_args+=(--mcphub-token-file /run/input/mcphub-token)
+fi
 while IFS= read -r config; do
-  sync_args+=(--hermes "$config")
+  sync_args+=(--hermes "/hermes/${config#"$HERMES_DATA_DIR"/}")
 done < <(find "$HERMES_DATA_DIR" -name config.yaml -type f -print)
-python3 "${STACK_DIR:-/opt/mike-ai/stack}/platform/mcp/sync-clients.py" "${sync_args[@]}"
+docker run --rm --entrypoint python \
+  -v "$STACK_DIR:/stack:ro" \
+  -v "$HERMES_DATA_DIR:/hermes:rw" \
+  "${token_mount[@]}" \
+  casaderoll/mcphub:1.1.0 \
+  /stack/platform/mcp/sync-clients.py "${sync_args[@]}"
 
-"${STACK_DIR:-/opt/mike-ai/stack}/platform/hermes/install-skills.sh"
+"$STACK_DIR/platform/hermes/install-skills.sh"
 docker exec "$HERMES_CONTAINER" hermes profile list
 printf 'HERMES_PROFILES_OK\n'
diff --git a/platform/hermes/install-skills.sh b/platform/hermes/install-skills.sh
index 829a467..a139694 100755
--- a/platform/hermes/install-skills.sh
+++ b/platform/hermes/install-skills.sh
@@ -1,31 +1,39 @@
 #!/usr/bin/env bash
 set -Eeuo pipefail
 
-STACK_DIR=${STACK_DIR:-/opt/mike-ai/stack}
-HERMES_DATA_DIR=${HERMES_DATA_DIR:-/data/hermes}
-SKILL_SOURCE=$STACK_DIR/platform/hermes/skills/athena-operator/SKILL.md
-PROFILES=(fast medium large ultra uncensored)
+STACK_DIR=${STACK_DIR:-$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)}
+HERMES_DATA_DIR=${HERMES_DATA_DIR:-/mnt/nvme-storage/appdata/Hermes-Agent}
+PROFILE_MATRIX=${PROFILE_MATRIX:-$STACK_DIR/config/profile-matrix.json}
+SKILL_ROOT=$STACK_DIR/platform/hermes/skills
 
 die() { printf 'FEHLER: %s\n' "$*" >&2; exit 1; }
 [[ $EUID -eq 0 ]] || die "Bitte als root ausführen."
-[[ -s $SKILL_SOURCE ]] || die "Skill-Quelle fehlt: $SKILL_SOURCE"
-grep -Fxq -- 'name: athena-operator' "$SKILL_SOURCE" || \
-  die "Skill-Quelle hat kein gültiges Athena-Operator-Frontmatter."
+[[ -s $PROFILE_MATRIX ]] || die "Profilmatrix fehlt: $PROFILE_MATRIX"
 
 install_skill() {
-  local root=$1 target=$1/platform/athena-operator/SKILL.md
+  local source=$1 root=$2 name target
+  name=${source%/SKILL.md}
+  name=${name##*/}
+  target=$root/platform/$name/SKILL.md
+  grep -Fxq -- "name: $name" "$source" || \
+    die "Skill-Quelle hat kein gültiges Frontmatter: $source"
   install -d -o 10000 -g 10000 -m 0750 "${target%/*}"
-  if [[ -s $target ]] && ! cmp -s "$SKILL_SOURCE" "$target"; then
+  if [[ -s $target ]] && ! cmp -s "$source" "$target"; then
     cp -a "$target" "$target.before-managed-update-$(date +%Y%m%d-%H%M%S)"
   fi
-  install -o 10000 -g 10000 -m 0640 "$SKILL_SOURCE" "$target"
-  cmp -s "$SKILL_SOURCE" "$target" || die "Skill-Synchronisierung fehlgeschlagen: $target"
+  install -o 10000 -g 10000 -m 0640 "$source" "$target"
+  cmp -s "$source" "$target" || die "Skill-Synchronisierung fehlgeschlagen: $target"
 }
 
-install_skill "$HERMES_DATA_DIR/skills"
-for profile in "${PROFILES[@]}"; do
-  [[ -d $HERMES_DATA_DIR/profiles/$profile ]] || continue
-  install_skill "$HERMES_DATA_DIR/profiles/$profile/skills"
+mapfile -t profiles < <(jq -r '.profiles[].id' "$PROFILE_MATRIX")
+
+for source in "$SKILL_ROOT"/*/SKILL.md; do
+  [[ -s $source ]] || continue
+  install_skill "$source" "$HERMES_DATA_DIR/skills"
+  for profile in "${profiles[@]}"; do
+    [[ -d $HERMES_DATA_DIR/profiles/$profile ]] || continue
+    install_skill "$source" "$HERMES_DATA_DIR/profiles/$profile/skills"
+  done
 done
 
-printf 'HERMES_ATHENA_OPERATOR_SKILL_OK\n'
+printf 'HERMES_SKILLS_OK\n'
diff --git a/platform/hermes/skills/mcphub-deployer/SKILL.md b/platform/hermes/skills/mcphub-deployer/SKILL.md
index 0147c03..6bef8fb 100644
--- a/platform/hermes/skills/mcphub-deployer/SKILL.md
+++ b/platform/hermes/skills/mcphub-deployer/SKILL.md
@@ -16,10 +16,10 @@ Use these paths directly. Do not search the filesystem for alternatives.
 - Host: Unraid `192.168.1.2`
 - Container: `MCPHub`
 - UI/base URL: `http://192.168.1.2:8787`
-- Operational source/build tree: `/mnt/nvme-storage/appdata/MCPHub/build`
+- Operational source/build tree: `/mnt/nvme-storage/appdata/MCPHub/build/repo`
 - Dockerfile: `platform/mcphub/Dockerfile` below that tree
-- Server declaration: `platform/mcphub/configure-settings.py`
-- Client registry: `config/mcp-registry.json`
+- Single server and client registry: `config/mcp-registry.json`
+- Registry renderer: `platform/mcphub/configure-settings.py` (normally unchanged)
 - Unraid template: `config/unraid-templates/my-MCPHub.xml`
 - Persistent state: `/mnt/nvme-storage/appdata/MCPHub`
 - Secrets: `/mnt/nvme-storage/appdata/MCPHub/secrets/<server>.env`, mode `0600`
@@ -91,8 +91,8 @@ Modify only the necessary fixed production files. Rules:
 - Put credentials only in the matching secret file with mode `0600`.
 - Never print, log, commit, summarize, or return a secret.
 - Use `/usr/local/bin/run-with-env` for secret-backed stdio servers.
-- Declare servers in `configure-settings.py`; do not manually treat
-  `mcp_settings.json` as the source of truth.
+- Declare servers only in `config/mcp-registry.json`; do not hard-code a server
+  in `configure-settings.py` and do not edit `mcp_settings.json` manually.
 - Preserve existing users, bearer keys, prompts, resources, enabled states,
   and per-tool toggles.
 - Build a new image tag. Never overwrite the tag currently running.
@@ -104,7 +104,7 @@ unfiltered large server to clients.
 
 ### 5. Deploy without collateral changes
 
-Build from `/mnt/nvme-storage/appdata/MCPHub/build`, then recreate only
+Build from `/mnt/nvme-storage/appdata/MCPHub/build/repo`, then recreate only
 `MCPHub` through Unraid DockerMan so it stays a managed Unraid container.
 Preserve all Appdata and mounts.
 
diff --git a/platform/mcp/install-tools.sh b/platform/mcp/install-tools.sh
index 32fa9d4..5bff534 100755
--- a/platform/mcp/install-tools.sh
+++ b/platform/mcp/install-tools.sh
@@ -11,13 +11,6 @@ if [[ -s $STACK_ENV ]]; then
   COMPOSE+=(--env-file "$STACK_ENV")
 fi
 COMPOSE+=(-f "$STACK_DIR/compose.yaml")
-export SEARXNG_SETTINGS_FILE="${SEARXNG_SETTINGS_FILE:-$MCP_DIR/../web-search/searxng-settings.yml}"
-
-[[ -s $SEARXNG_SETTINGS_FILE ]] || {
-  echo "SearXNG-Konfiguration fehlt: $SEARXNG_SETTINGS_FILE" >&2
-  exit 1
-}
-
 docker network inspect mike-ai-tools >/dev/null 2>&1 || \
   docker network create --internal --subnet 172.30.40.0/24 mike-ai-tools >/dev/null
 docker network inspect mike-ai-tools-egress >/dev/null 2>&1 || \
@@ -26,36 +19,11 @@ docker network inspect mike-ai-tools-egress >/dev/null 2>&1 || \
 # One administrative MCP exposes ATHENA.md plus the bounded host operator.
 "$MCP_DIR/../operator/install-operator.sh"
 
-services=(searxng tinysearch mcp-athena-operator)
+services=(mcp-athena-operator)
 # Portable Fach-MCPs laufen zentral im MCPHub auf Unraid. Ihre Compose-Blöcke
 # bleiben vorläufig als explizite Rollback-Profile erhalten, werden bei einer
 # normalen Athena-Installation aber weder gebaut noch gestartet.
 echo "ARR, Deemix, GitHub, Home Assistant und Navidrome werden über MCPHub auf Unraid bereitgestellt."
-# TinySearch keeps the embedding bundle outside the container. Download it
-# once on a fresh host; subsequent rebuilds reuse the named volume.
-#
-# Do not call `tinysearch setup` here. The pinned container image already
-# contains a complete Playwright/Chromium installation, while that command
-# unconditionally tries to install Chromium again. On IPv4-only hosts this
-# redundant download can hang indefinitely. Prepare only the persistent ONNX
-# bundle that is actually absent on a fresh installation.
-docker volume create mike-ai-tools_tinysearch-models >/dev/null
-tiny_image="marcellm01/tinysearch@sha256:7a7d0585f5000f462e699e42b97409715826a9e2edcd09a166afa93a4b7cba31"
-if ! docker run --rm --entrypoint test \
-  -v mike-ai-tools_tinysearch-models:/data/models "$tiny_image" \
-  -f /data/models/all-minilm-l6-v2-onnx/model.onnx; then
-  echo "TinySearch-Modell wird einmalig geladen."
-  docker run --rm --entrypoint python \
-    -v mike-ai-tools_tinysearch-models:/data/models "$tiny_image" \
-    -c 'from tinysearch.services.onnx_bundle_service import ensure_onnx_bundle_sync; ensure_onnx_bundle_sync("fast")'
-fi
-
 "${COMPOSE[@]}" up -d --build "${services[@]}"
 
-for webui in mike-ai-open-webui Open-WebUI; do
-  if docker container inspect "$webui" >/dev/null 2>&1; then
-    docker network connect mike-ai-tools "$webui" 2>/dev/null || true
-  fi
-done
-
 "${COMPOSE[@]}" ps
diff --git a/platform/mcp/sync-clients.py b/platform/mcp/sync-clients.py
index 90e36dc..a583901 100755
--- a/platform/mcp/sync-clients.py
+++ b/platform/mcp/sync-clients.py
@@ -13,6 +13,7 @@ import time
 
 BEGIN = "# BEGIN MANAGED MCP SERVERS"
 END = "# END MANAGED MCP SERVERS"
+CLIENT_TOKEN = ""
 
 
 def env_file(path: str) -> dict[str, str]:
@@ -37,7 +38,10 @@ def enabled(item: dict) -> bool:
     if source:
         values = env_file(source)
         url_ready = bool(item.get("url")) or bool(values.get(item.get("url_env", "")))
-        key_ready = not item.get("key_env") or bool(values.get(item["key_env"]))
+        key_ready = (not item.get("key_env")
+                     or bool(values.get(item["key_env"]))
+                     or (item.get("key_env") == "MCPHUB_BEARER_TOKEN"
+                         and bool(CLIENT_TOKEN)))
         return url_ready and key_ready
     return True
 
@@ -47,6 +51,8 @@ def resolved(item: dict) -> tuple[str, str]:
         values = env_file(item["env_file"])
         url = item.get("url") or values[item["url_env"]]
         key = values.get(item.get("key_env", ""), "")
+        if not key and item.get("key_env") == "MCPHUB_BEARER_TOKEN":
+            key = CLIENT_TOKEN
         return url, key
     return item["url"], ""
 
@@ -72,6 +78,16 @@ def hermes_block(items: list[dict]) -> str:
         ])
         if key:
             lines.extend(["    headers:", f"      Authorization: {yaml_quote('Bearer ' + key)}"])
+        if "tool_include" in item:
+            lines.append("    tools:")
+            lines.append("      include:")
+            for tool in item["tool_include"]:
+                lines.append(f"        - {yaml_quote(str(tool))}")
+        elif "tool_exclude" in item:
+            lines.append("    tools:")
+            lines.append("      exclude:")
+            for tool in item["tool_exclude"]:
+                lines.append(f"        - {yaml_quote(str(tool))}")
         lines.extend([
             f"    timeout: {int(item.get('timeout', 300))}",
             "    connect_timeout: 30",
@@ -103,6 +119,8 @@ def openwebui_connection(item: dict) -> dict:
     config = {"enable": True, "access_grants": []}
     if item.get("functions"):
         config["function_name_filter_list"] = item["functions"]
+    elif item.get("tool_include"):
+        config["function_name_filter_list"] = ",".join(item["tool_include"])
     return {
         "url": url, "path": "", "type": "mcp",
         "auth_type": item.get("auth_type", "none"), "headers": None,
@@ -135,11 +153,17 @@ def update_openwebui(db: pathlib.Path, items: list[dict]) -> None:
 
 
 def main() -> None:
+    global CLIENT_TOKEN
     parser = argparse.ArgumentParser()
     parser.add_argument("--registry", type=pathlib.Path, required=True)
     parser.add_argument("--hermes", type=pathlib.Path, action="append", default=[])
     parser.add_argument("--openwebui-db", type=pathlib.Path)
+    parser.add_argument("--mcphub-token-file", type=pathlib.Path)
     args = parser.parse_args()
+    if args.mcphub_token_file:
+        CLIENT_TOKEN = args.mcphub_token_file.read_text(encoding="utf-8").strip()
+        if not CLIENT_TOKEN:
+            raise SystemExit("MCPHub token file is empty")
     if args.hermes:
         block = hermes_block(active(args.registry, "hermes"))
         for path in args.hermes:
diff --git a/platform/mcphub/Dockerfile b/platform/mcphub/Dockerfile
index 5f36d25..97da1ef 100644
--- a/platform/mcphub/Dockerfile
+++ b/platform/mcphub/Dockerfile
@@ -2,11 +2,13 @@ FROM ghcr.io/github/github-mcp-server@sha256:1817b57d43916532dc002bdc5f344d639bd
 
 FROM ghcr.io/blakeem/navidrome-mcp:2.2.0@sha256:047f911a5a8f7cc8f185bb4d6e7ca6c435542edefff4694a00c2f718ab0ee7f5 AS navidrome
 
-FROM samanhappy/mcphub:1.0.32
+FROM samanhappy/mcphub:1.0.32@sha256:df34df85e639743d0bf4b64182fc2f47586bf544beb08c85a5e5d693891daad3
 
 ARG MCP_VERSION=1.29.0
 ARG ARR_MCP_VERSION=1.0.1
 ARG YT_DLP_VERSION=2026.7.4
+ARG FRITZ_MCP_VERSION=0.8.0
+ARG FRITZ_MCP_SHA256=47f4e2b5595a522aeda4aed9f5c65a57e500c3589c9f4262b8fbf72bd8c72bd4
 
 USER root
 
@@ -18,13 +20,21 @@ RUN python3 -m pip install --no-cache-dir \
       "arr-mcp[mcp]==${ARR_MCP_VERSION}" \
       "yt-dlp==${YT_DLP_VERSION}"
 
+# Released, checksum-pinned Go binary. Keeping the download in the reproducible
+# image build prevents MCPHub upgrades from silently dropping the Fritz tools.
+RUN mkdir -p /opt/casaderoll \
+    && python3 -c 'import sys,urllib.request; v=sys.argv[1]; urllib.request.urlretrieve(f"https://github.com/kambriso/fritzbox-mcp-server/releases/download/v{v}/fritz-mcp-linux-amd64", "/opt/casaderoll/fritz-mcp")' "${FRITZ_MCP_VERSION}" \
+    && echo "${FRITZ_MCP_SHA256}  /opt/casaderoll/fritz-mcp" | sha256sum -c - \
+    && chmod 0755 /opt/casaderoll/fritz-mcp
+
 COPY --from=github /server/github-mcp-server /usr/local/bin/github-mcp-server
 COPY --from=navidrome /app /opt/casaderoll/navidrome
 
 COPY platform/mcp/deemix_mcp.py /opt/casaderoll/mcps/deemix_mcp.py
-COPY platform/web-search/web_search_mcp.py /opt/casaderoll/mcps/web_search_mcp.py
 COPY platform/mcp/patches/mcp_sonarr.py /usr/local/lib/python3.13/site-packages/arr_mcp/mcp/mcp_sonarr.py
 COPY platform/mcp/patches/mcp_radarr.py /usr/local/lib/python3.13/site-packages/arr_mcp/mcp/mcp_radarr.py
+COPY config/mcp-registry.json /opt/casaderoll/config/mcp-registry.json
+COPY platform/mcphub/configure-settings.py /opt/casaderoll/configure-settings.py
 COPY platform/mcphub/run-with-env.py /usr/local/bin/run-with-env
 COPY platform/mcphub/casaderoll-entrypoint.sh /usr/local/bin/casaderoll-mcphub-entrypoint
 
@@ -36,7 +46,8 @@ RUN node -e 'const fs=require("node:fs"); const p="/opt/casaderoll/navidrome/dis
 
 ENV MCPHUB_SETTING_PATH=/app/data/ \
     REQUEST_TIMEOUT=120000 \
-    NODE_ENV=production
+    NODE_ENV=production \
+    FRITZ_MCP_VERSION=${FRITZ_MCP_VERSION}
 
 ENTRYPOINT ["/usr/local/bin/casaderoll-mcphub-entrypoint"]
 CMD ["/usr/local/bin/entrypoint.sh", "pnpm", "start"]
diff --git a/platform/mcphub/README.md b/platform/mcphub/README.md
index b33c9a8..9a5d2d3 100644
--- a/platform/mcphub/README.md
+++ b/platform/mcphub/README.md
@@ -9,8 +9,8 @@ sie sich einen Docker-Container und ein Appdata-Backup teilen.
 - ARR, Deemix, Navidrome und GitHub laufen als lokale stdio-Unterprozesse.
 - Home Assistant und MUA/Unraid sind vorhandene HTTP-MCP-Endpunkte und werden
   vom Hub direkt weitergereicht.
-- Der Web-Adapter kann hier laufen; SearXNG/TinySearch dürfen getrennte
-  Backend-Dienste bleiben.
+- Allgemeine Webrecherche bleibt ein eingebautes Hermes-Werkzeug. Der alte
+  Athena-Webadapter sowie SearXNG/TinySearch gehören nicht zum MCPHub-Image.
 - Athenas administrativer Operator ist hostgebunden und bleibt auf Athena.
   MCPHub reicht den vorhandenen, nur über WireGuard erreichbaren HTTP-Endpunkt
   `http://192.168.1.212:8202/mcp` als `/mcp/athena-operator` weiter. Dadurch
@@ -64,7 +64,7 @@ read-only-Aufruf prüfen und erst danach Clients auf
 `http://UNRAID-IP:8787/mcp/{server}` umstellen. Der alte Athena-Container wird
 erst gestoppt, wenn Hermes und OpenWebUI nachweislich über MCPHub funktionieren.
 
-Aktueller Stand: Athena Operator, ARR, Deemix, Navidrome, GitHub, Home Assistant
-und MUA/Unraid sind auf MCPHub registriert. Alte portable Athena-MCP-Container
-bleiben vorläufig als ausgeschaltetes Rückfallnetz bestehen. Der Web-Adapter
-ist der letzte noch offene Migrationspunkt.
+Aktueller Stand: Athena Operator, ARR, Deemix, Navidrome, GitHub, Home Assistant,
+MUA/Unraid und FRITZ!Box sind auf MCPHub registriert. Alte portable
+Athena-MCP-Container bleiben ausgeschaltet als kurzfristiges Rückfallnetz
+bestehen. Der frühere Webadapter ist nicht mehr Bestandteil des Images.
diff --git a/platform/mcphub/casaderoll-entrypoint.sh b/platform/mcphub/casaderoll-entrypoint.sh
index faa8090..69fca90 100644
--- a/platform/mcphub/casaderoll-entrypoint.sh
+++ b/platform/mcphub/casaderoll-entrypoint.sh
@@ -18,4 +18,13 @@ fi
 JWT_SECRET=$(cat "$jwt_file")
 export JWT_SECRET
 
+# Reconcile the persistent MCPHub state with the versioned registry on every
+# start. Existing users, bearer tokens and per-server enabled flags survive.
+# This makes image upgrades reproducible instead of relying on manual edits in
+# MCPHub's database/UI.
+settings_file="$state_dir/mcp_settings.json"
+python3 /opt/casaderoll/configure-settings.py \
+  "$settings_file" /run/secrets/mcphub \
+  --registry /opt/casaderoll/config/mcp-registry.json
+
 exec "$@"
diff --git a/platform/mcphub/configure-settings.py b/platform/mcphub/configure-settings.py
index d2f096c..0839563 100644
--- a/platform/mcphub/configure-settings.py
+++ b/platform/mcphub/configure-settings.py
@@ -11,6 +11,7 @@ import argparse
 import json
 import os
 import pathlib
+import re
 import secrets
 import tempfile
 import uuid
@@ -34,75 +35,70 @@ def env_file(path: pathlib.Path) -> dict[str, str]:
     return values
 
 
-def required(values: dict[str, str], key: str, source: pathlib.Path) -> str:
-    value = values.get(key, "").strip()
-    if not value:
-        raise SystemExit(f"{key} is missing in {source}")
+PLACEHOLDER = re.compile(r"\$\{([A-Za-z_][A-Za-z0-9_]*)\}")
+
+
+def expand(value: object, values: dict[str, str], source: pathlib.Path) -> object:
+    """Resolve secret placeholders without ever logging their values."""
+    if isinstance(value, str):
+        def replace(match: re.Match[str]) -> str:
+            key = match.group(1)
+            resolved = values.get(key, "").strip()
+            if not resolved:
+                raise SystemExit(f"{key} is missing in {source}")
+            return resolved
+        return PLACEHOLDER.sub(replace, value)
+    if isinstance(value, list):
+        return [expand(item, values, source) for item in value]
+    if isinstance(value, dict):
+        return {key: expand(item, values, source) for key, item in value.items()}
     return value
 
 
+def registry_servers(registry: pathlib.Path, secrets_dir: pathlib.Path,
+                     existing: dict[str, object]) -> dict[str, object]:
+    document = json.loads(registry.read_text(encoding="utf-8"))
+    if document.get("version") != 1:
+        raise SystemExit("Unsupported MCP registry schema")
+    result: dict[str, object] = {}
+    for item in document.get("servers", []):
+        spec = item.get("hub")
+        if not isinstance(spec, dict):
+            continue
+        server_id = str(item.get("hermes_id") or item["id"])
+        spec = dict(spec)
+        secret_name = str(spec.pop("secret_file", ""))
+        secret_path = secrets_dir / secret_name if secret_name else secrets_dir
+        values = env_file(secret_path) if secret_name else {}
+        rendered = expand(spec, values, secret_path)
+        if isinstance(rendered, dict) and isinstance(rendered.get("url"), str):
+            rendered["url"] = re.sub(r"(?<!:)//+", "/", rendered["url"])
+        previous = existing.get(server_id)
+        if isinstance(previous, dict) and "enabled" in previous:
+            rendered["enabled"] = bool(previous["enabled"])
+        result[server_id] = rendered
+    return result
+
+
 def main() -> None:
     parser = argparse.ArgumentParser()
     parser.add_argument("settings", type=pathlib.Path)
     parser.add_argument("secrets", type=pathlib.Path)
+    parser.add_argument(
+        "--registry", type=pathlib.Path,
+        default=pathlib.Path("/opt/casaderoll/config/mcp-registry.json"),
+    )
     parser.add_argument("--web-backend", default="", help="TinySearch MCP URL; empty keeps web disabled")
     parser.add_argument("--searxng", default="", help="SearXNG base URL")
     args = parser.parse_args()
 
+    args.settings.parent.mkdir(parents=True, exist_ok=True)
     settings = json.loads(args.settings.read_text(encoding="utf-8")) if args.settings.exists() else {}
-    ha_path = args.secrets / "homeassistant.env"
-    mua_path = args.secrets / "mua.env"
-    ha = env_file(ha_path)
-    mua = env_file(mua_path)
-
-    servers: dict[str, object] = {
-        "athena-operator": {
-            "type": "streamable-http",
-            "url": "http://192.168.1.212:8202/mcp",
-            "owner": "admin",
-            "enabled": True,
-        },
-        "arr": {
-            "type": "stdio",
-            "command": "/usr/local/bin/run-with-env",
-            "args": ["/run/secrets/mcphub/arr.env", "--", "arr-mcp", "--transport", "stdio", "--auth-type", "none"],
-            "enabled": True,
-        },
-        "deemix": {
-            "type": "stdio",
-            "command": "/usr/local/bin/run-with-env",
-            "args": ["/run/secrets/mcphub/deemix.env", "--", "python3", "/opt/casaderoll/mcps/deemix_mcp.py"],
-            "env": {"MCP_TRANSPORT": "stdio"},
-            "enabled": True,
-        },
-        "navidrome": {
-            "type": "stdio",
-            "command": "/usr/local/bin/run-with-env",
-            "args": ["/run/secrets/mcphub/navidrome.env", "--", "node", "/opt/casaderoll/navidrome/dist/index.js"],
-            "env": {"MCP_TRANSPORT": "stdio", "MCP_HTTP_EXPOSE": "false", "WEBUI_ENABLED": "false"},
-            "enabled": True,
-        },
-        "github": {
-            "type": "stdio",
-            "command": "/usr/local/bin/run-with-env",
-            "args": ["/run/secrets/mcphub/github.env", "--", "/usr/local/bin/github-mcp-server", "stdio", "--read-only", "--tools", "search_repositories,get_file_contents,search_code"],
-            "enabled": True,
-        },
-        "homeassistant": {
-            "type": "streamable-http",
-            "url": required(ha, "HASS_URL", ha_path).rstrip("/") + "/api/hass_mcp",
-            "headers": {"Authorization": "Bearer " + required(ha, "HASS_TOKEN", ha_path)},
-            "owner": "admin",
-            "enabled": True,
-        },
-        "unraid": {
-            "type": "streamable-http",
-            "url": required(mua, "MUA_MCP_URL", mua_path),
-            "headers": {"Authorization": "Bearer " + required(mua, "MUA_MCP_BEARER_TOKEN", mua_path)},
-            "owner": "admin",
-            "enabled": True,
-        },
-    }
+    servers = registry_servers(
+        args.registry,
+        args.secrets,
+        settings.get("mcpServers", {}) if isinstance(settings.get("mcpServers"), dict) else {},
+    )
     if args.web_backend:
         servers["web"] = {
             "type": "stdio",
@@ -144,7 +140,6 @@ def main() -> None:
     system = settings.setdefault("systemConfig", {})
     system.setdefault("routing", {})["skipAuth"] = False
 
-    args.settings.parent.mkdir(parents=True, exist_ok=True)
     fd, temporary = tempfile.mkstemp(prefix=".mcp-settings-", dir=args.settings.parent)
     try:
         with os.fdopen(fd, "w", encoding="utf-8") as handle:
diff --git a/platform/scripts/render-mcp-registry.py b/platform/scripts/render-mcp-registry.py
new file mode 100644
index 0000000..b837041
--- /dev/null
+++ b/platform/scripts/render-mcp-registry.py
@@ -0,0 +1,57 @@
+#!/usr/bin/env python3
+"""Render the human-readable MCP list from the single JSON registry."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import pathlib
+
+
+ROOT = pathlib.Path(__file__).resolve().parents[2]
+REGISTRY = ROOT / "config" / "mcp-registry.json"
+OUTPUT = ROOT / "docs" / "MCP_SERVERS.md"
+
+
+def render() -> str:
+    document = json.loads(REGISTRY.read_text(encoding="utf-8"))
+    rows = []
+    for item in document["servers"]:
+        if not item.get("hub"):
+            continue
+        clients = ", ".join(item.get("clients", [])) or "–"
+        selection = str(len(item["tool_include"])) if item.get("tool_include") else "alle"
+        rows.append(
+            f"| {item['name']} | `{item['hermes_id']}` | `{item['url']}` | "
+            f"{clients} | {selection} |"
+        )
+    return "\n".join([
+        "# MCP-Server",
+        "",
+        "Diese Datei wird aus `config/mcp-registry.json` erzeugt. Änderungen gehören nur in die JSON-Registry.",
+        "",
+        "| Server | Hermes-ID | Endpunkt | Clients | Werkzeuge |",
+        "|---|---|---|---|---:|",
+        *rows,
+        "",
+        "Allgemeine Webrecherche ist ein eingebautes Hermes-Werkzeug und kein MCPHub-Server.",
+        "",
+    ])
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--check", action="store_true")
+    args = parser.parse_args()
+    expected = render()
+    if args.check:
+        if not OUTPUT.is_file() or OUTPUT.read_text(encoding="utf-8") != expected:
+            raise SystemExit("MCP registry documentation is out of date")
+        print("MCP_REGISTRY_OK")
+        return
+    OUTPUT.write_text(expected, encoding="utf-8")
+    print("MCP_REGISTRY_RENDERED")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/platform/scripts/sync-profile-matrix.py b/platform/scripts/sync-profile-matrix.py
new file mode 100755
index 0000000..c93e31e
--- /dev/null
+++ b/platform/scripts/sync-profile-matrix.py
@@ -0,0 +1,126 @@
+#!/usr/bin/env python3
+"""Render and verify all client-facing profile data from one matrix."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import pathlib
+import re
+
+
+def load(path: pathlib.Path) -> dict:
+    data = json.loads(path.read_text(encoding="utf-8"))
+    if data.get("version") != 1 or not isinstance(data.get("profiles"), list):
+        raise SystemExit("Unsupported profile matrix schema")
+    ids: set[str] = set()
+    aliases: set[str] = set()
+    for item in data["profiles"]:
+        required = {"id", "alias", "context", "model_env", "description"}
+        missing = required - item.keys()
+        if missing:
+            raise SystemExit(f"Profile entry missing: {sorted(missing)}")
+        if item["id"] in ids or item["alias"] in aliases:
+            raise SystemExit(f"Duplicate profile or alias: {item['id']}")
+        if not isinstance(item["context"], int) or item["context"] < 8192:
+            raise SystemExit(f"Invalid context for {item['id']}")
+        ids.add(item["id"])
+        aliases.add(item["alias"])
+    if data.get("default_profile") not in ids:
+        raise SystemExit("default_profile is not defined")
+    return data
+
+
+def write_if_changed(path: pathlib.Path, content: str) -> None:
+    if path.exists() and path.read_text(encoding="utf-8") == content:
+        return
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(content, encoding="utf-8")
+
+
+def render_router(data: dict) -> str:
+    payload = {
+        "profiles": {
+            item["id"]: {
+                "context": item["context"],
+                "model_alias": item["alias"],
+            }
+            for item in data["profiles"]
+        }
+    }
+    return json.dumps(payload, indent=2, ensure_ascii=False) + "\n"
+
+
+def render_docs(data: dict) -> str:
+    lines = [
+        "# Profilmatrix",
+        "",
+        "Diese Datei wird aus `config/profile-matrix.json` erzeugt. Änderungen gehören nur in die JSON-Matrix.",
+        "",
+        f"Standardprofil: **{data['default_profile']}** · globales Ausgabelimit: **{data['max_output_tokens']} Token**",
+        "",
+        "| Profil | API-Alias | Kontext | Modell | GPU-Verteilung | Vision | MTP |",
+        "|---|---|---:|---|---|---|---:|",
+    ]
+    for item in data["profiles"]:
+        lines.append(
+            f"| {item['id']} | `{item['alias']}` | {item['context']:,} | "
+            f"{item.get('model_family', '')} | {item.get('gpu_split', '')} | "
+            f"{'ja' if item.get('vision') else 'nein'} | {item.get('mtp', '')} |"
+        )
+    lines.extend(["", "## Zweck", ""])
+    for item in data["profiles"]:
+        lines.append(f"- **{item['id']}**: {item['description']}")
+    return "\n".join(lines) + "\n"
+
+
+def verify_compose(data: dict, compose: pathlib.Path) -> None:
+    text = compose.read_text(encoding="utf-8")
+    expected_cap = f'MAX_GENERATION_TOKENS: "{int(data["max_output_tokens"])}"'
+    if expected_cap not in text:
+        raise SystemExit("Router output-token cap drift")
+    for item in data["profiles"]:
+        block_match = re.search(
+            rf"(?ms)^  llama-{re.escape(item['id'])}:\n(?P<body>.*?)(?=^  [a-zA-Z0-9_-]+:|\Z)",
+            text,
+        )
+        if not block_match:
+            raise SystemExit(f"Compose service llama-{item['id']} is missing")
+        block = block_match.group("body")
+        if f"- {item['alias']}" not in block:
+            raise SystemExit(f"Compose alias drift for {item['id']}")
+        env_prefix = item["id"].upper()
+        if f"${{{env_prefix}_CONTEXT:-{item['context']}}}" not in block:
+            raise SystemExit(f"Compose context drift for {item['id']}")
+
+
+def main() -> None:
+    root = pathlib.Path(__file__).resolve().parents[2]
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--matrix", type=pathlib.Path,
+                        default=root / "config/profile-matrix.json")
+    parser.add_argument("--router", type=pathlib.Path,
+                        default=root / "router/router_profiles.json")
+    parser.add_argument("--docs", type=pathlib.Path,
+                        default=root / "docs/STANDARD_PROFILE_MATRIX.md")
+    parser.add_argument("--compose", type=pathlib.Path,
+                        default=root / "compose.yaml")
+    parser.add_argument("--check", action="store_true")
+    args = parser.parse_args()
+    data = load(args.matrix)
+    router = render_router(data)
+    docs = render_docs(data)
+    verify_compose(data, args.compose)
+    if args.check:
+        if args.router.read_text(encoding="utf-8") != router:
+            raise SystemExit("router_profiles.json is not generated from the matrix")
+        if args.docs.read_text(encoding="utf-8") != docs:
+            raise SystemExit("profile documentation is not generated from the matrix")
+    else:
+        write_if_changed(args.router, router)
+        write_if_changed(args.docs, docs)
+    print("PROFILE_MATRIX_OK")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/restore.sh b/restore.sh
index 1c1f64a..9128008 100755
--- a/restore.sh
+++ b/restore.sh
@@ -23,8 +23,7 @@ tar -xzf "$ARCHIVE" -C "$work"
 [[ -d $work/volumes ]] || die "Backup enthält keine Docker-Volumes."
 
 # Stop only users of the restored volumes. WireGuard, SSH and networking stay up.
-for container in mike-ai-open-webui mike-ai-router mike-ai-profile-controller \
-  mike-ai-piper mike-ai-tools-tinysearch; do
+for container in mike-ai-router mike-ai-profile-controller mike-ai-piper; do
   if [[ $(docker inspect -f '{{.State.Running}}' "$container" 2>/dev/null || true) == true ]]; then
     docker stop "$container" >/dev/null
   fi
@@ -42,19 +41,11 @@ restore_volume() {
   rsync -a --delete "$source/" "$mountpoint/"
 }
 
-restore_volume mike-ai_open-webui-data "$work/volumes/open-webui-data"
 restore_volume mike-ai_piper-data "$work/volumes/piper-data"
 restore_volume mike-ai_router-state "$work/volumes/router-state"
 restore_volume mike-ai_router-images "$work/volumes/router-images"
-restore_volume mike-ai-tools_tinysearch-models "$work/volumes/tinysearch-models"
 
 cd "$ROOT_DIR"
-docker compose --env-file /etc/mike-ai/stack.env \
-  --profile homeassistant --profile arr --profile navidrome --profile deemix --profile github \
-  up -d
-python3 platform/mcp/sync-clients.py \
-  --registry config/mcp-registry.json \
-  --hermes /data/hermes/config.yaml \
-  --openwebui-db "$(docker volume inspect -f '{{.Mountpoint}}' mike-ai_open-webui-data)/webui.db"
+./manage.sh deploy core
 
 printf 'ATHENA_RESTORE_OK %s\n' "$ARCHIVE"
diff --git a/router/ai_profile_router.py b/router/ai_profile_router.py
index ae3b849..698f093 100755
--- a/router/ai_profile_router.py
+++ b/router/ai_profile_router.py
@@ -62,6 +62,7 @@ import logging
 import os
 import queue
 import re
+import select
 import socket
 import subprocess
 import sys
@@ -2098,6 +2099,21 @@ class Handler(BaseHTTPRequestHandler):
                 chunk = resp.read1(16384)
                 if not chunk:
                     break
+                # A closed downstream socket may still accept one small write
+                # into the kernel buffer. Detect the peer FIN before writing so
+                # slow upstream streams are cancelled promptly and do not keep
+                # an inference lease occupied until generation completes.
+                try:
+                    readable, _, _ = select.select([self.connection], [], [], 0)
+                    if readable:
+                        flags = socket.MSG_PEEK | getattr(socket, "MSG_DONTWAIT", 0)
+                        if self.connection.recv(1, flags) == b"":
+                            log.info("Client trennte Streaming-Verbindung; Upstream wird abgebrochen")
+                            break
+                except (BlockingIOError, InterruptedError):
+                    pass
+                except OSError:
+                    break
                 self.wfile.write(chunk)
                 self.wfile.flush()
         except (OSError, http.client.HTTPException) as e:
diff --git a/smoke-test.sh b/smoke-test.sh
new file mode 100755
index 0000000..6d22e8a
--- /dev/null
+++ b/smoke-test.sh
@@ -0,0 +1,82 @@
+#!/usr/bin/env bash
+# Short, read-only end-to-end test for every Athena release.
+set -Eeuo pipefail
+
+ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+ENV_FILE=${STACK_ENV:-/etc/mike-ai/stack.env}
+
+pass() { printf 'PASS  %s\n' "$*"; }
+fail() { printf 'FAIL  %s\n' "$*" >&2; exit 1; }
+
+[[ $EUID -eq 0 ]] || fail "Bitte als root ausfuehren."
+command -v docker >/dev/null || fail "Docker fehlt."
+[[ -s $ENV_FILE ]] || fail "Stack-Konfiguration fehlt: $ENV_FILE"
+
+cd "$ROOT_DIR"
+./manage.sh validate >/dev/null
+pass "Profilmatrix, MCP-Liste und Compose-Konfiguration stimmen"
+
+healthy() {
+  local name=$1 state health
+  state=$(docker inspect -f '{{.State.Status}}' "$name" 2>/dev/null || true)
+  health=$(docker inspect -f '{{if .State.Health}}{{.State.Health.Status}}{{end}}' "$name" 2>/dev/null || true)
+  [[ $state == running && ( -z $health || $health == healthy ) ]]
+}
+
+for name in \
+  mike-ai-wireguard-gateway \
+  mike-ai-profile-controller \
+  mike-ai-router \
+  mike-ai-piper \
+  mike-ai-xtts \
+  mike-ai-tts-gateway \
+  mike-ai-mcp-athena-operator \
+  mike-ai-backup; do
+  healthy "$name" || fail "$name fehlt oder ist nicht gesund"
+done
+pass "Athena-Kerndienste sind gesund"
+
+active_llama=$(docker ps --format '{{.Names}}' | \
+  grep -Ec '^mike-ai-llama-(fast|medium|large|ultra|uncensored)$' || true)
+[[ $active_llama -eq 1 ]] || fail "$active_llama aktive llama-Profile (erwartet: 1)"
+pass "Genau ein llama.cpp-Profil ist aktiv"
+
+for legacy in \
+  mike-ai-open-webui \
+  mike-ai-hermes \
+  mike-ai-hermes-webui \
+  mike-ai-hermes-webui-vpn-proxy \
+  mike-ai-mcp-web \
+  mike-ai-tools-searxng \
+  mike-ai-tools-tinysearch \
+  mike-ai-mcp-platform-context; do
+  [[ $(docker inspect -f '{{.State.Running}}' "$legacy" 2>/dev/null || true) != true ]] || \
+    fail "Altlast laeuft noch: $legacy"
+done
+pass "OpenWebUI, alte Hermes-Dienste und migrierte MCPs sind aus"
+
+docker exec mike-ai-router python - <<'PY' >/dev/null
+import json
+import os
+import urllib.request
+
+key = os.environ["ROUTER_API_KEY"]
+for path in ("/health", "/v1/models"):
+    request = urllib.request.Request(
+        "http://127.0.0.1:8081" + path,
+        headers={"Authorization": "Bearer " + key},
+    )
+    with urllib.request.urlopen(request, timeout=10) as response:
+        body = response.read()
+        if response.status != 200:
+            raise SystemExit(f"{path}: HTTP {response.status}")
+        if path.endswith("models") and not json.loads(body).get("data"):
+            raise SystemExit("Router meldet keine Modelle")
+PY
+pass "Router-Liveness und OpenAI-Modellliste funktionieren"
+
+latest=/data/docker-backups/athena-latest.tar.gz
+[[ -s $latest ]] || fail "Kein gueltiges Athena-Backup: $latest"
+pass "Automatisches Athena-Backup ist vorhanden"
+
+printf '\nATHENA_E2E_OK\n'