fast-agent-mcp 0.2.14__py3-none-any.whl → 0.2.17__py3-none-any.whl

mcp_agent/llm/augmented_llm_playback.py

@@ -1,9 +1,11 @@
-from typing import Any, List
+from typing import Any, List, Type
 
+from mcp_agent.core.exceptions import ModelConfigError
 from mcp_agent.core.prompt import Prompt
 from mcp_agent.llm.augmented_llm import RequestParams
 from mcp_agent.llm.augmented_llm_passthrough import PassthroughLLM
 from mcp_agent.llm.provider_types import Provider
+from mcp_agent.mcp.interfaces import ModelT
 from mcp_agent.mcp.prompt_message_multipart import PromptMessageMultipart
 from mcp_agent.mcp.prompts.prompt_helpers import MessageContent
 
@@ -82,3 +84,21 @@ class PlaybackLLM(PassthroughLLM):
         )
 
         return response
+
+    async def structured(
+        self,
+        multipart_messages: List[PromptMessageMultipart],
+        model: Type[ModelT],
+        request_params: RequestParams | None = None,
+    ) -> tuple[ModelT | None, PromptMessageMultipart]:
+        """
+        Handle structured requests by returning the next assistant message.
+        """
+
+        if -1 == self._current_index:
+            raise ModelConfigError("Use generate() to load playback history")
+
+        return self._structured_from_multipart(
+            self._get_next_assistant_message(),
+            model,
+        )

mcp_agent/llm/memory.py

@@ -19,7 +19,7 @@ class Memory(Protocol, Generic[MessageParamT]):
 
     def append(self, message: MessageParamT, is_prompt: bool = False) -> None: ...
 
-    def get(self, include_history: bool = True) -> List[MessageParamT]: ...
+    def get(self, include_completion_history: bool = True) -> List[MessageParamT]: ...
 
     def clear(self, clear_prompts: bool = False) -> None: ...
 
@@ -75,7 +75,7 @@ class SimpleMemory(Memory, Generic[MessageParamT]):
         else:
             self.history.append(message)
 
-    def get(self, include_history: bool = True) -> List[MessageParamT]:
+    def get(self, include_completion_history: bool = True) -> List[MessageParamT]:
         """
         Get all messages in memory.
 
@@ -86,7 +86,7 @@ class SimpleMemory(Memory, Generic[MessageParamT]):
         Returns:
             Combined list of prompt messages and optionally history messages
         """
-        if include_history:
+        if include_completion_history:
             return self.prompt_messages + self.history
         else:
             return self.prompt_messages.copy()

mcp_agent/llm/model_factory.py

@@ -12,6 +12,7 @@ from mcp_agent.llm.provider_types import Provider
 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,
     }
 
@@ -161,7 +163,7 @@ class ModelFactory:
         Creates a factory function that follows the attach_llm protocol.
 
         Args:
-            model_string: The model specification string (e.g. "gpt-4o.high")
+            model_string: The model specification string (e.g. "gpt-4.1")
             request_params: Optional parameters to configure LLM behavior
 
         Returns:

mcp_agent/llm/provider_key_manager.py

@@ -14,6 +14,7 @@ PROVIDER_ENVIRONMENT_MAP: Dict[str, str] = {
     "anthropic": "ANTHROPIC_API_KEY",
     "openai": "OPENAI_API_KEY",
     "deepseek": "DEEPSEEK_API_KEY",
+    "google": "GOOGLE_API_KEY",
     "openrouter": "OPENROUTER_API_KEY",
     "generic": "GENERIC_API_KEY",
 }

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"

mcp_agent/llm/providers/augmented_llm_anthropic.py

@@ -1,4 +1,4 @@
-from typing import TYPE_CHECKING, List
+from typing import TYPE_CHECKING, List, Tuple, Type
 
 from mcp.types import EmbeddedResource, ImageContent, TextContent
 
@@ -10,6 +10,7 @@ from mcp_agent.llm.providers.multipart_converter_anthropic import (
 from mcp_agent.llm.providers.sampling_converter_anthropic import (
     AnthropicSamplingConverter,
 )
+from mcp_agent.mcp.interfaces import ModelT
 from mcp_agent.mcp.prompt_message_multipart import PromptMessageMultipart
 
 if TYPE_CHECKING:
@@ -50,6 +51,19 @@ class AnthropicAugmentedLLM(AugmentedLLM[MessageParam, Message]):
     selecting appropriate tools, and determining what information to retain.
     """
 
+    # Anthropic-specific parameter exclusions
+    ANTHROPIC_EXCLUDE_FIELDS = {
+        AugmentedLLM.PARAM_MESSAGES,
+        AugmentedLLM.PARAM_MODEL,
+        AugmentedLLM.PARAM_SYSTEM_PROMPT,
+        AugmentedLLM.PARAM_STOP_SEQUENCES,
+        AugmentedLLM.PARAM_MAX_TOKENS,
+        AugmentedLLM.PARAM_METADATA,
+        AugmentedLLM.PARAM_USE_HISTORY,
+        AugmentedLLM.PARAM_MAX_ITERATIONS,
+        AugmentedLLM.PARAM_PARALLEL_TOOL_CALLS,
+    }
+
     def __init__(self, *args, **kwargs) -> None:
         # Initialize logger - keep it simple without name reference
         self.logger = get_logger(__name__)
@@ -73,7 +87,7 @@ class AnthropicAugmentedLLM(AugmentedLLM[MessageParam, Message]):
         assert self.context.config
         return self.context.config.anthropic.base_url if self.context.config.anthropic else None
 
-    async def generate_internal(
+    async def _anthropic_completion(
         self,
         message_param,
         request_params: RequestParams | None = None,
@@ -100,7 +114,7 @@ class AnthropicAugmentedLLM(AugmentedLLM[MessageParam, Message]):
 
         # Always include prompt messages, but only include conversation history
         # if use_history is True
-        messages.extend(self.history.get(include_history=params.use_history))
+        messages.extend(self.history.get(include_completion_history=params.use_history))
 
         messages.append(message_param)
 
@@ -120,7 +134,8 @@ class AnthropicAugmentedLLM(AugmentedLLM[MessageParam, Message]):
 
         for i in range(params.max_iterations):
             self._log_chat_progress(self.chat_turn(), model=model)
-            arguments = {
+            # Create base arguments dictionary
+            base_args = {
                 "model": model,
                 "messages": messages,
                 "system": self.instruction or params.systemPrompt,
@@ -129,10 +144,12 @@ class AnthropicAugmentedLLM(AugmentedLLM[MessageParam, Message]):
             }
 
             if params.maxTokens is not None:
-                arguments["max_tokens"] = params.maxTokens
+                base_args["max_tokens"] = params.maxTokens
 
-            if params.metadata:
-                arguments = {**arguments, **params.metadata}
+            # Use the base class method to prepare all arguments with Anthropic-specific exclusions
+            arguments = self.prepare_provider_arguments(
+                base_args, params, self.ANTHROPIC_EXCLUDE_FIELDS
+            )
 
             self.logger.debug(f"{arguments}")
 
@@ -265,7 +282,7 @@ class AnthropicAugmentedLLM(AugmentedLLM[MessageParam, Message]):
         # Keep the prompt messages separate
         if params.use_history:
             # Get current prompt messages
-            prompt_messages = self.history.get(include_history=False)
+            prompt_messages = self.history.get(include_completion_history=False)
 
             # Calculate new conversation messages (excluding prompts)
             new_messages = messages[len(prompt_messages) :]
@@ -288,7 +305,7 @@ class AnthropicAugmentedLLM(AugmentedLLM[MessageParam, Message]):
         Override this method to use a different LLM.
 
         """
-        res = await self.generate_internal(
+        res = await self._anthropic_completion(
             message_param=message_param,
             request_params=request_params,
         )
@@ -298,6 +315,7 @@ class AnthropicAugmentedLLM(AugmentedLLM[MessageParam, Message]):
         self,
         multipart_messages: List["PromptMessageMultipart"],
         request_params: RequestParams | None = None,
+        is_template: bool = False,
     ) -> PromptMessageMultipart:
         # Check the last message role
         last_message = multipart_messages[-1]
@@ -310,7 +328,7 @@ class AnthropicAugmentedLLM(AugmentedLLM[MessageParam, Message]):
         for msg in messages_to_add:
             converted.append(AnthropicConverter.convert_to_anthropic(msg))
 
-        self.history.extend(converted, is_prompt=True)
+        self.history.extend(converted, is_prompt=is_template)
 
         if last_message.role == "user":
             self.logger.debug("Last message in prompt is from user, generating assistant response")
@@ -321,6 +339,27 @@ class AnthropicAugmentedLLM(AugmentedLLM[MessageParam, Message]):
             self.logger.debug("Last message in prompt is from assistant, returning it directly")
             return last_message
 
+    async def _apply_prompt_provider_specific_structured(
+        self,
+        multipart_messages: List[PromptMessageMultipart],
+        model: Type[ModelT],
+        request_params: RequestParams | None = None,
+    ) -> Tuple[ModelT | None, PromptMessageMultipart]:  # noqa: F821
+        request_params = self.get_request_params(request_params)
+
+        multipart_messages[-1].add_text(
+            """YOU MUST RESPOND IN THE FOLLOWING FORMAT:
+            {schema}
+            RESPOND ONLY WITH THE JSON, NO PREAMBLE OR CODE FENCES """.format(
+                schema=model.model_json_schema()
+            )
+        )
+
+        result: PromptMessageMultipart = await self._apply_prompt_provider_specific(
+            multipart_messages, request_params
+        )
+        return self._structured_from_multipart(result, model)
+
     @classmethod
     def convert_message_to_message_param(cls, message: Message, **kwargs) -> MessageParam:
         """Convert a response object to an input parameter object to allow LLM calls to be chained."""

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 @@ DEFAULT_DEEPSEEK_MODEL = "deepseekchat"  # current Deepseek only has two type mo
 
 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

mcp_agent/llm/providers/augmented_llm_generic.py

@@ -10,8 +10,10 @@ DEFAULT_OLLAMA_API_KEY = "ollama"
 
 
 class GenericAugmentedLLM(OpenAIAugmentedLLM):
-    def __init__(self, *args, provider=Provider.GENERIC, **kwargs) -> None:
-        super().__init__(*args, **kwargs)  # Properly pass args and kwargs to parent
+    def __init__(self, *args, **kwargs) -> None:
+        super().__init__(
+            *args, provider=Provider.GENERIC, **kwargs
+        )  # Properly pass args and kwargs to parent
 
     def _initialize_default_params(self, kwargs: dict) -> RequestParams:
         """Initialize Generic  parameters"""

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