diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index b26ddaf1cf..0a8a86136c 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -5015,6 +5015,15 @@ class LlamaCppBackend:
                 _accumulated_predicted_ms += _it.get("predicted_ms", 0)
                 _accumulated_predicted_n += _it.get("predicted_n", 0)
 
+                # When the caller opted out of parallel tool calls
+                # (parallel_tool_calls=False), enforce at most one call
+                # per assistant turn even if llama-server emitted more.
+                # llama.cpp's parallel_tool_calls flag isn't enforced by
+                # every jinja template (see ggml-org/llama.cpp#22043),
+                # so this client-side cap is the only guarantee.
+                if parallel_tool_calls is False and tool_calls:
+                    tool_calls = tool_calls[:1]
+
                 assistant_msg = {"role": "assistant", "content": content_text}
                 if tool_calls:
                     assistant_msg["tool_calls"] = tool_calls
diff --git a/studio/backend/tests/test_sampling_params_routing.py b/studio/backend/tests/test_sampling_params_routing.py
index 431ff4cd65..0143b7770d 100644
--- a/studio/backend/tests/test_sampling_params_routing.py
+++ b/studio/backend/tests/test_sampling_params_routing.py
@@ -876,6 +876,26 @@ def test_local_anthropic_disable_parallel_tool_use_translation():
     assert _extract({"type": "auto", "disable_parallel_tool_use": "yes"}) is None
 
 
+def test_gguf_tool_loop_enforces_parallel_tool_calls_false():
+    """llama.cpp's `parallel_tool_calls` flag is not enforced by every
+    jinja template (see ggml-org/llama.cpp#22043), so when the caller
+    opted out we must cap tool_calls to the first entry before the
+    agentic loop executes them. The cap is a single-line slice in
+    `generate_chat_completion_with_tools`; pin the contract."""
+    from pathlib import Path
+
+    src = Path(__file__).resolve().parent.parent / "core/inference/llama_cpp.py"
+    text = src.read_text()
+    assert "if parallel_tool_calls is False and tool_calls" in text, (
+        "GGUF tool loop must enforce parallel_tool_calls=False by "
+        "truncating tool_calls before assistant_msg is built; that "
+        "is the client-side guarantee llama-server's flag does not "
+        "give us. See routes/inference.py and chat-adapter.ts for "
+        "the wire-side forwarding of the same flag."
+    )
+    assert "tool_calls = tool_calls[:1]" in text
+
+
 def test_local_anthropic_passthrough_helpers_accept_parallel_tool_calls():
     """The Anthropic-compat client-tool passthrough helpers
     (`_anthropic_passthrough_stream` /
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index 6f01c0becd..97533c4f7d 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -1529,15 +1529,19 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
             // Optional sampling extensions; local llama-server already
             // accepts `stop` / `seed` / `frequency_penalty` via
             // _build_passthrough_payload (routes/inference.py:4884) and
-            // silently ignores fields it does not recognise.
+            // silently ignores fields it does not recognise. llama-server
+            // documents `parallel_tool_calls` defaulting to FALSE
+            // (https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md);
+            // forward the user's preference unconditionally so the
+            // default-on UI state actually enables parallel tool calls
+            // there. External providers default to true everywhere; the
+            // external branch above keeps its opt-in-on-false shape.
             ...(params.frequencyPenalty !== 0
               ? { frequency_penalty: params.frequencyPenalty }
               : {}),
             ...(params.seed !== null ? { seed: params.seed } : {}),
             ...(params.stop.length > 0 ? { stop: params.stop } : {}),
-            ...(params.parallelToolCalls === false
-              ? { parallel_tool_calls: false }
-              : {}),
+            parallel_tool_calls: params.parallelToolCalls,
             image_base64: imageBase64,
             audio_base64: audioBase64,
             cancel_id: cancelId,