diff --git a/cmd/mavend/voicewire.go b/cmd/mavend/voicewire.go
index f3a2f7f..f476128 100644
--- a/cmd/mavend/voicewire.go
+++ b/cmd/mavend/voicewire.go
@@ -386,8 +386,13 @@ func modelSeam(cfg *config.Config, resident *llm.Client) (router.Completer, *llm
 		return resident, nil
 	}
 	ws := cfg.Workstation
+	remote := llm.New(ws.URL, time.Duration(ws.Timeout))
+	remote.SetToken(ws.Token)
+	if ws.Token == "" {
+		log.Printf("voice: no workstation.token — mavgpud refuses an unauthenticated request, so this reads as a card that is always busy")
+	}
 	pair := llm.NewPair(
-		llm.New(ws.URL, time.Duration(ws.Timeout)),
+		remote,
 		resident,
 		ws.Health,
 		time.Duration(ws.Probe),
diff --git a/deploy/mavend.json b/deploy/mavend.json
index aa0d95b..349d163 100644
--- a/deploy/mavend.json
+++ b/deploy/mavend.json
@@ -94,6 +94,7 @@
   ],
   "workstation": {
     "url": "http://192.168.1.105:8080",
+    "token": "${MAVEN_GPU_TOKEN}",
     "probe": "15s",
     "timeout": "90s",
     "stt": {
diff --git a/deploy/telegram.env.example b/deploy/telegram.env.example
index 8542bcf..3717da7 100644
--- a/deploy/telegram.env.example
+++ b/deploy/telegram.env.example
@@ -10,3 +10,12 @@ TELEGRAM_CHAT_ID=
 #   ntfy token add --expires=never maven
 # Read access is not needed — mavend publishes and never subscribes.
 NTFY_TOKEN=
+
+# Bearer token for mavgpud, the workstation's GPU supervisor (V-673). It fronts
+# the big model on a LAN port, so the token is the whole boundary in front of
+# the card. Any long random string; mint one with:
+#   openssl rand -hex 32
+# The same value goes in a file on workpc, named by token_file in
+# deploy/mavgpud.json. Unset here and every workstation turn falls back to the
+# resident model, because mavgpud answers 401 and Maven reads that as down.
+MAVEN_GPU_TOKEN=
diff --git a/internal/config/workstation.go b/internal/config/workstation.go
index 6fab688..6b34d2a 100644
--- a/internal/config/workstation.go
+++ b/internal/config/workstation.go
@@ -27,6 +27,13 @@ type WorkstationConfig struct {
 	// signal, so it must be the supervisor's endpoint and not llama-server's.
 	Health string `json:"health,omitempty"`
 
+	// Token — the bearer credential mavgpud requires, expanded from the
+	// environment like every other secret here. It is what stops anything on
+	// the LAN spending the card, so a URL that is not loopback needs one.
+	// Wrong or missing reads as a workstation that is down, and Maven falls
+	// back to the resident model.
+	Token string `json:"token,omitempty"`
+
 	// Probe — how often admission is re-checked. 0 ⇒ DefaultWorkstationProbe.
 	// Nothing on the hot path waits for it: the answer is cached and read
 	// atomically, so this only sets how late Maven notices the card came back.
diff --git a/internal/llm/client.go b/internal/llm/client.go
index 4d77ae6..bc93b36 100644
--- a/internal/llm/client.go
+++ b/internal/llm/client.go
@@ -51,6 +51,10 @@ type Client struct {
 	base string
 	swap SwapGate
 	http *http.Client
+	// token — the bearer credential for a server that asks for one. Empty for
+	// the resident model, which is reached over loopback on the same box.
+	// mavgpud on the workstation requires it: that hop is on the LAN.
+	token string
 
 	// gate / background — priority on the single llama-server slot. Set once
 	// at wiring time (SetGate), read on every request. nil gate ⇒ no gating,
@@ -101,6 +105,27 @@ func New(baseURL string, timeout time.Duration) *Client {
 	return &Client{base: baseURL, http: &http.Client{Timeout: timeout}}
 }
 
+// SetToken installs the bearer credential this client sends. Wiring-time, like
+// SetGate: an empty token means the server is not asking for one.
+func (c *Client) SetToken(t string) {
+	c.mu.Lock()
+	c.token = t
+	c.mu.Unlock()
+}
+
+// authorize adds the credential when there is one. Exported to the package so
+// the Pair prober signs /health with the same token as the completion — a
+// probe that answers 401 would otherwise read as a workstation that is down,
+// and Maven would fall back forever without saying why.
+func (c *Client) authorize(req *http.Request) {
+	c.mu.RLock()
+	t := c.token
+	c.mu.RUnlock()
+	if t != "" {
+		req.Header.Set("Authorization", "Bearer "+t)
+	}
+}
+
 // SetBaseURL re-points the client at another llama-server. Safe to call while
 // requests are in flight: a request that already read the old base finishes
 // against the old base (or fails, and every caller of Complete has a fallback),
@@ -196,6 +221,7 @@ func (c *Client) Complete(ctx context.Context, r Req) (string, error) {
 		return "", err
 	}
 	req.Header.Set("Content-Type", "application/json")
+	c.authorize(req)
 	httpResp, err := c.http.Do(req)
 	if err != nil {
 		return "", err
diff --git a/internal/llm/remote.go b/internal/llm/remote.go
index 30b1143..69b19ba 100644
--- a/internal/llm/remote.go
+++ b/internal/llm/remote.go
@@ -132,6 +132,9 @@ func (p *Pair) probe(ctx context.Context) {
 		p.set(false)
 		return
 	}
+	if p.remote != nil {
+		p.remote.authorize(req)
+	}
 	resp, err := p.http.Do(req)
 	if err != nil {
 		p.set(false)
diff --git a/internal/llm/remote_test.go b/internal/llm/remote_test.go
index 3169b42..787b627 100644
--- a/internal/llm/remote_test.go
+++ b/internal/llm/remote_test.go
@@ -250,3 +250,35 @@ func TestNoFloorIsAnError(t *testing.T) {
 		t.Fatalf("err = %v, want ErrNoFloor", err)
 	}
 }
+
+// The workstation is behind mavgpud, which requires a bearer token on the
+// completion and on /health alike. A probe that did not carry it would answer
+// 401, Pair would read that as a card that is busy, and every turn would fall
+// back to the resident model with nothing in the log naming why.
+func TestPairSignsTheProbeAndTheCompletion(t *testing.T) {
+	seen := make(chan string, 2)
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		seen <- r.Header.Get("Authorization")
+		if r.URL.Path == "/health" {
+			return
+		}
+		_, _ = w.Write([]byte(`{"choices":[{"message":{"content":"ok"}}]}`))
+	}))
+	defer srv.Close()
+
+	remote := New(srv.URL, time.Second)
+	remote.SetToken("s3cret")
+	p := NewPair(remote, New(srv.URL, time.Second), srv.URL+"/health", time.Hour)
+	p.probe(context.Background())
+	if !p.Available() {
+		t.Fatal("the probe did not admit an answering workstation")
+	}
+	if _, err := p.Complete(context.Background(), Req{User: "привет"}); err != nil {
+		t.Fatalf("complete: %v", err)
+	}
+	for i := 0; i < 2; i++ {
+		if got := <-seen; got != "Bearer s3cret" {
+			t.Errorf("request %d carried %q, want the bearer token", i, got)
+		}
+	}
+}