diff --git a/docs/concepts/vendors.md b/docs/concepts/vendors.md
index e214b5d..b82c2d7 100644
--- a/docs/concepts/vendors.md
+++ b/docs/concepts/vendors.md
@@ -24,7 +24,7 @@ Used with `agent.with_llm()` for the cascading flow (ASR → LLM → TTS).
 | `OpenAI` | OpenAI | `model` for Agora-managed global models; `api_key`, `base_url`, `model` for BYOK |
 | `AzureOpenAI` | Azure OpenAI | `api_key`, `model`, `endpoint`, `deployment_name` |
 | `Anthropic` | Anthropic | `api_key`, `model`, `url`, `headers`, `max_tokens` |
-| `Gemini` | Google Gemini | `api_key`, `model` |
+| `Gemini` | Google Gemini | `api_key`, `model`; optional `url` |
 | `Groq` | Groq | `api_key`, `model`, `base_url` |
 | `VertexAILLM` | Google Vertex AI | `api_key`, `model`, `project_id`, `location` |
 | `AmazonBedrock` | Amazon Bedrock | `access_key`, `secret_key`, `region`, `model` |
diff --git a/docs/reference/vendors.md b/docs/reference/vendors.md
index d1d96c2..697109e 100644
--- a/docs/reference/vendors.md
+++ b/docs/reference/vendors.md
@@ -176,6 +176,7 @@ llm = Anthropic(
 |---|---|---|---|---|
 | `api_key` | `str` | Yes | — | Google AI API key |
 | `model` | `str` | Yes | — | Model name |
+| `url` | `str` | No | `None` | Custom Gemini `streamGenerateContent` URL. When omitted, the SDK constructs one from `model` and `api_key`. |
 | `temperature` | `float` | No | `None` | Sampling temperature (0.0–2.0) |
 | `top_p` | `float` | No | `None` | Nucleus sampling (0.0–1.0) |
 | `top_k` | `int` | No | `None` | Top-k sampling |
@@ -194,7 +195,10 @@ llm = Anthropic(
 ```python
 from agora_agent import Gemini
 
-llm = Gemini(api_key='your-google-key', model='gemini-2.0-flash-exp')
+llm = Gemini(
+    api_key='your-google-key',
+    model='gemini-2.0-flash-exp',
+)
 ```
 
 ### Other LLM vendors
diff --git a/src/agora_agent/agentkit/vendors/llm.py b/src/agora_agent/agentkit/vendors/llm.py
index bfcb201..3e31992 100644
--- a/src/agora_agent/agentkit/vendors/llm.py
+++ b/src/agora_agent/agentkit/vendors/llm.py
@@ -304,8 +304,10 @@ def to_config(self) -> Dict[str, Any]:
             params["max_output_tokens"] = self.options.max_output_tokens
 
         config: Dict[str, Any] = {
-            "url": self.options.url or "https://generativelanguage.googleapis.com/v1beta/models",
-            "api_key": self.options.api_key,
+            "url": self.options.url or (
+                f"https://generativelanguage.googleapis.com/v1beta/models/"
+                f"{self.options.model}:streamGenerateContent?alt=sse&key={self.options.api_key}"
+            ),
             "params": params,
             "style": "gemini",
             "input_modalities": self.options.input_modalities or ["text"],
@@ -394,7 +396,9 @@ def to_config(self) -> Dict[str, Any]:
                 f"{self.options.project_id}/locations/{self.options.location}/"
                 f"publishers/google/models/{self.options.model}:streamGenerateContent?alt=sse"
             )
-        return Gemini(**options).to_config()
+        config = Gemini(**options).to_config()
+        config["api_key"] = self.options.api_key
+        return config
 
 
 class AmazonBedrockOptions(BaseModel):
diff --git a/tests/custom/test_llm_vendors.py b/tests/custom/test_llm_vendors.py
index 0d07cf4..b9cb90a 100644
--- a/tests/custom/test_llm_vendors.py
+++ b/tests/custom/test_llm_vendors.py
@@ -69,6 +69,32 @@ def test_vertex_ai_llm_includes_project_routing() -> None:
     assert "location" not in config.get("params", {})
 
 
+def test_vertex_ai_llm_preserves_explicit_url() -> None:
+    config = VertexAILLM(
+        api_key="vertex-token",
+        model="gemini-2.0-flash",
+        project_id="project",
+        location="us-central1",
+        url="https://vertex.example.com/custom-endpoint",
+    ).to_config()
+
+    assert config["url"] == "https://vertex.example.com/custom-endpoint"
+    assert config["api_key"] == "vertex-token"
+    assert config["params"]["model"] == "gemini-2.0-flash"
+
+
+def test_gemini_constructs_url_from_api_key_and_model() -> None:
+    config = Gemini(
+        api_key="google-key",
+        model="gemini-2.0-flash",
+    ).to_config()
+
+    assert config["url"] == "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash:streamGenerateContent?alt=sse&key=google-key"
+    assert "api_key" not in config
+    assert config["style"] == "gemini"
+    assert config["params"]["model"] == "gemini-2.0-flash"
+
+
 def test_amazon_bedrock_serializes_as_bedrock_style() -> None:
     config = AmazonBedrock(
         access_key="aws-access",
diff --git a/tests/custom/test_request_body.py b/tests/custom/test_request_body.py
index 083ed0a..689dbfa 100644
--- a/tests/custom/test_request_body.py
+++ b/tests/custom/test_request_body.py
@@ -898,9 +898,15 @@ def test_byok_anthropic_llm_params() -> None:
 
 
 def test_byok_gemini_llm_params() -> None:
-    agent = Agent(test_client()).with_llm(Gemini(api_key="gemini-key", model="gemini-2.0-flash"))
+    agent = Agent(test_client()).with_llm(
+        Gemini(
+            api_key="gemini-key",
+            model="gemini-2.0-flash",
+        )
+    )
     props = build_properties(agent, allow_missing={"asr", "tts"})
-    assert props["llm"]["api_key"] == "gemini-key"
+    assert "api_key" not in props["llm"]
+    assert props["llm"]["url"] == "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash:streamGenerateContent?alt=sse&key=gemini-key"
     assert props["llm"]["style"] == "gemini"
     assert props["llm"]["params"]["model"] == "gemini-2.0-flash"