google gemini (oai) support (#95)

Author: evalstate

src/mcp_agent/config.py

@@ -139,6 +139,19 @@ class DeepSeekSettings(BaseModel):
     model_config = ConfigDict(extra="allow", arbitrary_types_allowed=True)
 
 
+class GoogleSettings(BaseModel):
+    """
+    Settings for using OpenAI models in the fast-agent application.
+    """
+
+    api_key: str | None = None
+    # reasoning_effort: Literal["low", "medium", "high"] = "medium"
+
+    base_url: str | None = None
+
+    model_config = ConfigDict(extra="allow", arbitrary_types_allowed=True)
+
+
 class GenericSettings(BaseModel):
     """
     Settings for using OpenAI models in the fast-agent application.
@@ -264,6 +277,9 @@ class Settings(BaseSettings):
     deepseek: DeepSeekSettings | None = None
     """Settings for using DeepSeek models in the fast-agent application"""
 
+    google: GoogleSettings | None = None
+    """Settings for using DeepSeek models in the fast-agent application"""
+
     openrouter: OpenRouterSettings | None = None
     """Settings for using OpenRouter models in the fast-agent application"""
 

src/mcp_agent/llm/model_factory.py

@@ -12,6 +12,7 @@
 from mcp_agent.llm.providers.augmented_llm_anthropic import AnthropicAugmentedLLM
 from mcp_agent.llm.providers.augmented_llm_deepseek import DeepSeekAugmentedLLM
 from mcp_agent.llm.providers.augmented_llm_generic import GenericAugmentedLLM
+from mcp_agent.llm.providers.augmented_llm_google import GoogleAugmentedLLM
 from mcp_agent.llm.providers.augmented_llm_openai import OpenAIAugmentedLLM
 from mcp_agent.llm.providers.augmented_llm_openrouter import OpenRouterAugmentedLLM
 from mcp_agent.mcp.interfaces import AugmentedLLMProtocol
@@ -107,6 +108,7 @@ class ModelFactory:
         Provider.FAST_AGENT: PassthroughLLM,
         Provider.DEEPSEEK: DeepSeekAugmentedLLM,
         Provider.GENERIC: GenericAugmentedLLM,
+        Provider.GOOGLE: GoogleAugmentedLLM,  # type: ignore
         Provider.OPENROUTER: OpenRouterAugmentedLLM,
     }
 

src/mcp_agent/llm/provider_key_manager.py

@@ -14,6 +14,7 @@
     "anthropic": "ANTHROPIC_API_KEY",
     "openai": "OPENAI_API_KEY",
     "deepseek": "DEEPSEEK_API_KEY",
+    "google": "GOOGLE_API_KEY",
     "openrouter": "OPENROUTER_API_KEY",
     "generic": "GENERIC_API_KEY",
 }

src/mcp_agent/llm/provider_types.py

@@ -11,6 +11,7 @@ class Provider(Enum):
     ANTHROPIC = "anthropic"
     OPENAI = "openai"
     FAST_AGENT = "fast-agent"
+    GOOGLE = "google"
     DEEPSEEK = "deepseek"
     GENERIC = "generic"
-    OPENROUTER = "openrouter"
+    OPENROUTER = "openrouter"

src/mcp_agent/llm/providers/augmented_llm_deepseek.py

@@ -1,4 +1,3 @@
-
 from mcp_agent.core.request_params import RequestParams
 from mcp_agent.llm.provider_types import Provider
 from mcp_agent.llm.providers.augmented_llm_openai import OpenAIAugmentedLLM
@@ -9,7 +8,6 @@
 
 class DeepSeekAugmentedLLM(OpenAIAugmentedLLM):
     def __init__(self, *args, **kwargs) -> None:
-        kwargs["provider_name"] = "Deepseek"  # Set provider name in kwargs
         super().__init__(
             *args, provider=Provider.DEEPSEEK, **kwargs
         )  # Properly pass args and kwargs to parent

src/mcp_agent/llm/providers/augmented_llm_google.py

@@ -0,0 +1,30 @@
+from mcp_agent.core.request_params import RequestParams
+from mcp_agent.llm.provider_types import Provider
+from mcp_agent.llm.providers.augmented_llm_openai import OpenAIAugmentedLLM
+
+GOOGLE_BASE_URL = "https://generativelanguage.googleapis.com/v1beta/openai"
+DEFAULT_GOOGLE_MODEL = "gemini-2.0-flash"
+
+
+class GoogleAugmentedLLM(OpenAIAugmentedLLM):
+    def __init__(self, *args, **kwargs) -> None:
+        super().__init__(*args, provider=Provider.GOOGLE, **kwargs)
+
+    def _initialize_default_params(self, kwargs: dict) -> RequestParams:
+        """Initialize Google OpenAI Compatibility default parameters"""
+        chosen_model = kwargs.get("model", DEFAULT_GOOGLE_MODEL)
+
+        return RequestParams(
+            model=chosen_model,
+            systemPrompt=self.instruction,
+            parallel_tool_calls=False,
+            max_iterations=10,
+            use_history=True,
+        )
+
+    def _base_url(self) -> str:
+        base_url = None
+        if self.context.config and self.context.config.google:
+            base_url = self.context.config.google.base_url
+
+        return base_url if base_url else GOOGLE_BASE_URL

tests/e2e/smoke/test_e2e_smoke.py

@@ -27,6 +27,7 @@
         #        "generic.qwen2.5:latest",
         "generic.llama3.2:latest",
         "openrouter.google/gemini-2.0-flash-001",
+        "google.gemini-2.0-flash",
     ],
 )
 async def test_basic_textual_prompting(fast_agent, model_name):
@@ -260,6 +261,7 @@ async def agent_function():
         "gpt-4.1",
         "gpt-4.1-nano",
         "gpt-4.1-mini",
+        "google.gemini-2.0-flash",
     ],
 )
 async def test_basic_tool_calling(fast_agent, model_name):

tests/e2e/structured/test_structured_outputs.py

@@ -125,6 +125,7 @@ async def create_structured():
         # "haiku", -- anthropic do not support structured outputs this way
         "gpt-4.1-mini",
         "openrouter.google/gemini-2.0-flash-001",
+        "google.gemini-2.0-flash",
     ],
 )
 async def test_structured_output_with_response_format_overriden(fast_agent, model_name):
@@ -156,10 +157,7 @@ async def create_structured():
 @pytest.mark.e2e
 @pytest.mark.parametrize(
     "model_name",
-    [
-        "gpt-4.1-mini",
-        "haiku",
-    ],
+    ["gpt-4.1-mini", "haiku", "google.gemini-2.0-flash"],
 )
 async def test_history_management_with_structured(fast_agent, model_name):
     """Test that the agent can generate structured response with response_format_specified."""