letta-nightly 0.7.20.dev20250521104258__py3-none-any.whl → 0.7.21.dev20250522104246__py3-none-any.whl

letta/agents/letta_agent_batch.py

@@ -145,7 +145,7 @@ class LettaAgentBatch(BaseAgent):
         agent_mapping = {
             agent_state.id: agent_state
             for agent_state in await self.agent_manager.get_agents_by_ids_async(
-                agent_ids=[request.agent_id for request in batch_requests], actor=self.actor
+                agent_ids=[request.agent_id for request in batch_requests], include_relationships=["tools", "memory"], actor=self.actor
             )
         }
 
@@ -267,64 +267,121 @@ class LettaAgentBatch(BaseAgent):
 
     @trace_method
     async def _collect_resume_context(self, llm_batch_id: str) -> _ResumeContext:
-        # NOTE: We only continue for items with successful results
-        batch_items = await self.batch_manager.list_llm_batch_items_async(llm_batch_id=llm_batch_id, request_status=JobStatus.completed)
-
-        agent_ids = []
-        provider_results = {}
-        request_status_updates: List[RequestStatusUpdateInfo] = []
+        """
+        Collect context for resuming operations from completed batch items.
 
-        for item in batch_items:
-            aid = item.agent_id
-            agent_ids.append(aid)
-            provider_results[aid] = item.batch_request_result.result
+        Args:
+            llm_batch_id: The ID of the batch to collect context for
 
-        agent_states = await self.agent_manager.get_agents_by_ids_async(agent_ids, actor=self.actor)
-        agent_state_map = {agent.id: agent for agent in agent_states}
+        Returns:
+            _ResumeContext object containing all necessary data for resumption
+        """
+        # Fetch only completed batch items
+        batch_items = await self.batch_manager.list_llm_batch_items_async(llm_batch_id=llm_batch_id, request_status=JobStatus.completed)
 
-        name_map, args_map, cont_map = {}, {}, {}
-        for aid in agent_ids:
-            # status bookkeeping
-            pr = provider_results[aid]
-            status = (
-                JobStatus.completed
-                if isinstance(pr, BetaMessageBatchSucceededResult)
-                else (
-                    JobStatus.failed
-                    if isinstance(pr, BetaMessageBatchErroredResult)
-                    else JobStatus.cancelled if isinstance(pr, BetaMessageBatchCanceledResult) else JobStatus.expired
-                )
+        # Exit early if no items to process
+        if not batch_items:
+            return _ResumeContext(
+                batch_items=[],
+                agent_ids=[],
+                agent_state_map={},
+                provider_results={},
+                tool_call_name_map={},
+                tool_call_args_map={},
+                should_continue_map={},
+                request_status_updates=[],
             )
-            request_status_updates.append(RequestStatusUpdateInfo(llm_batch_id=llm_batch_id, agent_id=aid, request_status=status))
 
-            # translate provider‑specific response → OpenAI‑style tool call (unchanged)
-            llm_client = LLMClient.create(
-                provider_type=item.llm_config.model_endpoint_type,
-                put_inner_thoughts_first=True,
-                actor=self.actor,
-            )
-            tool_call = (
-                llm_client.convert_response_to_chat_completion(
-                    response_data=pr.message.model_dump(), input_messages=[], llm_config=item.llm_config
-                )
-                .choices[0]
-                .message.tool_calls[0]
-            )
+        # Extract agent IDs and organize items by agent ID
+        agent_ids = [item.agent_id for item in batch_items]
+        batch_item_map = {item.agent_id: item for item in batch_items}
 
-            name, args, cont = self._extract_tool_call_and_decide_continue(tool_call, item.step_state)
-            name_map[aid], args_map[aid], cont_map[aid] = name, args, cont
+        # Collect provider results
+        provider_results = {item.agent_id: item.batch_request_result.result for item in batch_items}
+
+        # Fetch agent states in a single call
+        agent_states = await self.agent_manager.get_agents_by_ids_async(
+            agent_ids=agent_ids, include_relationships=["tools", "memory"], actor=self.actor
+        )
+        agent_state_map = {agent.id: agent for agent in agent_states}
+
+        # Process each agent's results
+        tool_call_results = self._process_agent_results(
+            agent_ids=agent_ids, batch_item_map=batch_item_map, provider_results=provider_results, llm_batch_id=llm_batch_id
+        )
 
         return _ResumeContext(
             batch_items=batch_items,
             agent_ids=agent_ids,
             agent_state_map=agent_state_map,
             provider_results=provider_results,
-            tool_call_name_map=name_map,
-            tool_call_args_map=args_map,
-            should_continue_map=cont_map,
-            request_status_updates=request_status_updates,
+            tool_call_name_map=tool_call_results.name_map,
+            tool_call_args_map=tool_call_results.args_map,
+            should_continue_map=tool_call_results.cont_map,
+            request_status_updates=tool_call_results.status_updates,
+        )
+
+    def _process_agent_results(self, agent_ids, batch_item_map, provider_results, llm_batch_id):
+        """
+        Process the results for each agent, extracting tool calls and determining continuation status.
+
+        Returns:
+            A namedtuple containing name_map, args_map, cont_map, and status_updates
+        """
+        from collections import namedtuple
+
+        ToolCallResults = namedtuple("ToolCallResults", ["name_map", "args_map", "cont_map", "status_updates"])
+
+        name_map, args_map, cont_map = {}, {}, {}
+        request_status_updates = []
+
+        for aid in agent_ids:
+            item = batch_item_map[aid]
+            result = provider_results[aid]
+
+            # Determine job status based on result type
+            status = self._determine_job_status(result)
+            request_status_updates.append(RequestStatusUpdateInfo(llm_batch_id=llm_batch_id, agent_id=aid, request_status=status))
+
+            # Process tool calls
+            name, args, cont = self._extract_tool_call_from_result(item, result)
+            name_map[aid], args_map[aid], cont_map[aid] = name, args, cont
+
+        return ToolCallResults(name_map, args_map, cont_map, request_status_updates)
+
+    def _determine_job_status(self, result):
+        """Determine job status based on result type"""
+        if isinstance(result, BetaMessageBatchSucceededResult):
+            return JobStatus.completed
+        elif isinstance(result, BetaMessageBatchErroredResult):
+            return JobStatus.failed
+        elif isinstance(result, BetaMessageBatchCanceledResult):
+            return JobStatus.cancelled
+        else:
+            return JobStatus.expired
+
+    def _extract_tool_call_from_result(self, item, result):
+        """Extract tool call information from a result"""
+        llm_client = LLMClient.create(
+            provider_type=item.llm_config.model_endpoint_type,
+            put_inner_thoughts_first=True,
+            actor=self.actor,
+        )
+
+        # If result isn't a successful type, we can't extract a tool call
+        if not isinstance(result, BetaMessageBatchSucceededResult):
+            return None, None, False
+
+        tool_call = (
+            llm_client.convert_response_to_chat_completion(
+                response_data=result.message.model_dump(), input_messages=[], llm_config=item.llm_config
+            )
+            .choices[0]
+            .message.tool_calls[0]
         )
 
+        return self._extract_tool_call_and_decide_continue(tool_call, item.step_state)
+
     def _update_request_statuses(self, updates: List[RequestStatusUpdateInfo]) -> None:
         if updates:
             self.batch_manager.bulk_update_llm_batch_items_request_status_by_agent(updates=updates)
@@ -556,16 +613,6 @@ class LettaAgentBatch(BaseAgent):
         in_context_messages = await self._rebuild_memory_async(current_in_context_messages + new_in_context_messages, agent_state)
         return in_context_messages
 
-    # TODO: Make this a bullk function
-    def _rebuild_memory(
-        self,
-        in_context_messages: List[Message],
-        agent_state: AgentState,
-        num_messages: int | None = None,
-        num_archival_memories: int | None = None,
-    ) -> List[Message]:
-        return super()._rebuild_memory(in_context_messages, agent_state)
-
     # Not used in batch.
     async def step(self, input_messages: List[MessageCreate], max_steps: int = 10) -> LettaResponse:
         raise NotImplementedError

letta/agents/voice_agent.py

@@ -154,7 +154,7 @@ class VoiceAgent(BaseAgent):
         # TODO: Define max steps here
         for _ in range(max_steps):
             # Rebuild memory each loop
-            in_context_messages = self._rebuild_memory(in_context_messages, agent_state)
+            in_context_messages = await self._rebuild_memory_async(in_context_messages, agent_state)
             openai_messages = convert_in_context_letta_messages_to_openai(in_context_messages, exclude_system_messages=True)
             openai_messages.extend(in_memory_message_history)
 
@@ -292,14 +292,14 @@ class VoiceAgent(BaseAgent):
             agent_id=self.agent_id, message_ids=[m.id for m in new_in_context_messages], actor=self.actor
         )
 
-    def _rebuild_memory(
+    async def _rebuild_memory_async(
         self,
         in_context_messages: List[Message],
         agent_state: AgentState,
         num_messages: int | None = None,
         num_archival_memories: int | None = None,
     ) -> List[Message]:
-        return super()._rebuild_memory(
+        return await super()._rebuild_memory_async(
             in_context_messages, agent_state, num_messages=self.num_messages, num_archival_memories=self.num_archival_memories
         )
 
@@ -438,7 +438,7 @@ class VoiceAgent(BaseAgent):
         if start_date and end_date and start_date > end_date:
             start_date, end_date = end_date, start_date
 
-        archival_results = self.agent_manager.list_passages(
+        archival_results = await self.agent_manager.list_passages_async(
             actor=self.actor,
             agent_id=self.agent_id,
             query_text=archival_query,
@@ -457,7 +457,7 @@ class VoiceAgent(BaseAgent):
         keyword_results = {}
         if convo_keyword_queries:
             for keyword in convo_keyword_queries:
-                messages = self.message_manager.list_messages_for_agent(
+                messages = await self.message_manager.list_messages_for_agent_async(
                     agent_id=self.agent_id,
                     actor=self.actor,
                     query_text=keyword,

letta/client/client.py

@@ -2773,11 +2773,8 @@ class LocalClient(AbstractClient):
 
     # humans / personas
 
-    def get_block_id(self, name: str, label: str) -> str:
-        block = self.server.block_manager.get_blocks(actor=self.user, template_name=name, label=label, is_template=True)
-        if not block:
-            return None
-        return block[0].id
+    def get_block_id(self, name: str, label: str) -> str | None:
+        return None
 
     def create_human(self, name: str, text: str):
         """
@@ -2812,7 +2809,7 @@ class LocalClient(AbstractClient):
         Returns:
             humans (List[Human]): List of human blocks
         """
-        return self.server.block_manager.get_blocks(actor=self.user, label="human", is_template=True)
+        return []
 
     def list_personas(self) -> List[Persona]:
         """
@@ -2821,7 +2818,7 @@ class LocalClient(AbstractClient):
         Returns:
             personas (List[Persona]): List of persona blocks
         """
-        return self.server.block_manager.get_blocks(actor=self.user, label="persona", is_template=True)
+        return []
 
     def update_human(self, human_id: str, text: str):
         """
@@ -2879,7 +2876,7 @@ class LocalClient(AbstractClient):
         assert id, f"Human ID must be provided"
         return Human(**self.server.block_manager.get_block_by_id(id, actor=self.user).model_dump())
 
-    def get_persona_id(self, name: str) -> str:
+    def get_persona_id(self, name: str) -> str | None:
         """
         Get the ID of a persona block template
 
@@ -2889,12 +2886,9 @@ class LocalClient(AbstractClient):
         Returns:
             id (str): ID of the persona block
         """
-        persona = self.server.block_manager.get_blocks(actor=self.user, template_name=name, label="persona", is_template=True)
-        if not persona:
-            return None
-        return persona[0].id
+        return None
 
-    def get_human_id(self, name: str) -> str:
+    def get_human_id(self, name: str) -> str | None:
         """
         Get the ID of a human block template
 
@@ -2904,10 +2898,7 @@ class LocalClient(AbstractClient):
         Returns:
             id (str): ID of the human block
         """
-        human = self.server.block_manager.get_blocks(actor=self.user, template_name=name, label="human", is_template=True)
-        if not human:
-            return None
-        return human[0].id
+        return None
 
     def delete_persona(self, id: str):
         """
@@ -3381,7 +3372,7 @@ class LocalClient(AbstractClient):
         Returns:
             blocks (List[Block]): List of blocks
         """
-        return self.server.block_manager.get_blocks(actor=self.user, label=label, is_template=templates_only)
+        return []
 
     def create_block(
         self, label: str, value: str, limit: Optional[int] = None, template_name: Optional[str] = None, is_template: bool = False

letta/constants.py

@@ -19,6 +19,7 @@ MCP_TOOL_TAG_NAME_PREFIX = "mcp"  # full format, mcp:server_name
 LETTA_CORE_TOOL_MODULE_NAME = "letta.functions.function_sets.base"
 LETTA_MULTI_AGENT_TOOL_MODULE_NAME = "letta.functions.function_sets.multi_agent"
 LETTA_VOICE_TOOL_MODULE_NAME = "letta.functions.function_sets.voice"
+LETTA_BUILTIN_TOOL_MODULE_NAME = "letta.functions.function_sets.builtin"
 
 
 # String in the error message for when the context window is too large
@@ -83,9 +84,19 @@ BASE_VOICE_SLEEPTIME_TOOLS = [
 ]
 # Multi agent tools
 MULTI_AGENT_TOOLS = ["send_message_to_agent_and_wait_for_reply", "send_message_to_agents_matching_tags", "send_message_to_agent_async"]
+
+# Built in tools
+BUILTIN_TOOLS = ["run_code", "web_search"]
+
 # Set of all built-in Letta tools
 LETTA_TOOL_SET = set(
-    BASE_TOOLS + BASE_MEMORY_TOOLS + MULTI_AGENT_TOOLS + BASE_SLEEPTIME_TOOLS + BASE_VOICE_SLEEPTIME_TOOLS + BASE_VOICE_SLEEPTIME_CHAT_TOOLS
+    BASE_TOOLS
+    + BASE_MEMORY_TOOLS
+    + MULTI_AGENT_TOOLS
+    + BASE_SLEEPTIME_TOOLS
+    + BASE_VOICE_SLEEPTIME_TOOLS
+    + BASE_VOICE_SLEEPTIME_CHAT_TOOLS
+    + BUILTIN_TOOLS
 )
 
 # The name of the tool used to send message to the user
@@ -179,6 +190,45 @@ LLM_MAX_TOKENS = {
     "gpt-3.5-turbo-0613": 4096,  # legacy
     "gpt-3.5-turbo-16k-0613": 16385,  # legacy
     "gpt-3.5-turbo-0301": 4096,  # legacy
+    "gemini-1.0-pro-vision-latest": 12288,
+    "gemini-pro-vision": 12288,
+    "gemini-1.5-pro-latest": 2000000,
+    "gemini-1.5-pro-001": 2000000,
+    "gemini-1.5-pro-002": 2000000,
+    "gemini-1.5-pro": 2000000,
+    "gemini-1.5-flash-latest": 1000000,
+    "gemini-1.5-flash-001": 1000000,
+    "gemini-1.5-flash-001-tuning": 16384,
+    "gemini-1.5-flash": 1000000,
+    "gemini-1.5-flash-002": 1000000,
+    "gemini-1.5-flash-8b": 1000000,
+    "gemini-1.5-flash-8b-001": 1000000,
+    "gemini-1.5-flash-8b-latest": 1000000,
+    "gemini-1.5-flash-8b-exp-0827": 1000000,
+    "gemini-1.5-flash-8b-exp-0924": 1000000,
+    "gemini-2.5-pro-exp-03-25": 1048576,
+    "gemini-2.5-pro-preview-03-25": 1048576,
+    "gemini-2.5-flash-preview-04-17": 1048576,
+    "gemini-2.5-flash-preview-05-20": 1048576,
+    "gemini-2.5-flash-preview-04-17-thinking": 1048576,
+    "gemini-2.5-pro-preview-05-06": 1048576,
+    "gemini-2.0-flash-exp": 1048576,
+    "gemini-2.0-flash": 1048576,
+    "gemini-2.0-flash-001": 1048576,
+    "gemini-2.0-flash-exp-image-generation": 1048576,
+    "gemini-2.0-flash-lite-001": 1048576,
+    "gemini-2.0-flash-lite": 1048576,
+    "gemini-2.0-flash-preview-image-generation": 32768,
+    "gemini-2.0-flash-lite-preview-02-05": 1048576,
+    "gemini-2.0-flash-lite-preview": 1048576,
+    "gemini-2.0-pro-exp": 1048576,
+    "gemini-2.0-pro-exp-02-05": 1048576,
+    "gemini-exp-1206": 1048576,
+    "gemini-2.0-flash-thinking-exp-01-21": 1048576,
+    "gemini-2.0-flash-thinking-exp": 1048576,
+    "gemini-2.0-flash-thinking-exp-1219": 1048576,
+    "gemini-2.5-flash-preview-tts": 32768,
+    "gemini-2.5-pro-preview-tts": 65536,
 }
 # The error message that Letta will receive
 # MESSAGE_SUMMARY_WARNING_STR = f"Warning: the conversation history will soon reach its maximum length and be trimmed. Make sure to save any important information from the conversation to your memory before it is removed."
@@ -230,3 +280,7 @@ RETRIEVAL_QUERY_DEFAULT_PAGE_SIZE = 5
 
 MAX_FILENAME_LENGTH = 255
 RESERVED_FILENAMES = {"CON", "PRN", "AUX", "NUL", "COM1", "COM2", "LPT1", "LPT2"}
+
+WEB_SEARCH_CLIP_CONTENT = False
+WEB_SEARCH_INCLUDE_SCORE = False
+WEB_SEARCH_SEPARATOR = "\n" + "-" * 40 + "\n"

letta/functions/function_sets/builtin.py

@@ -0,0 +1,27 @@
+from typing import Literal
+
+
+async def web_search(query: str) -> str:
+    """
+    Search the web for information.
+    Args:
+        query (str): The query to search the web for.
+    Returns:
+        str: The search results.
+    """
+
+    raise NotImplementedError("This is only available on the latest agent architecture. Please contact the Letta team.")
+
+
+def run_code(code: str, language: Literal["python", "js", "ts", "r", "java"]) -> str:
+    """
+    Run code in a sandbox. Supports Python, Javascript, Typescript, R, and Java.
+
+    Args:
+        code (str): The code to run.
+        language (Literal["python", "js", "ts", "r", "java"]): The language of the code.
+    Returns:
+        str: The output of the code, the stdout, the stderr, and error traces (if any).
+    """
+
+    raise NotImplementedError("This is only available on the latest agent architecture. Please contact the Letta team.")

letta/groups/sleeptime_multi_agent_v2.py

@@ -190,7 +190,7 @@ class SleeptimeMultiAgentV2(BaseAgent):
             prior_messages = []
             if self.group.sleeptime_agent_frequency:
                 try:
-                    prior_messages = self.message_manager.list_messages_for_agent(
+                    prior_messages = await self.message_manager.list_messages_for_agent_async(
                         agent_id=foreground_agent_id,
                         actor=self.actor,
                         after=last_processed_message_id,

letta/interfaces/anthropic_streaming_interface.py

@@ -1,3 +1,4 @@
+import json
 from datetime import datetime, timezone
 from enum import Enum
 from typing import AsyncGenerator, List, Union
@@ -74,6 +75,7 @@ class AnthropicStreamingInterface:
         # usage trackers
         self.input_tokens = 0
         self.output_tokens = 0
+        self.model = None
 
         # reasoning object trackers
         self.reasoning_messages = []
@@ -88,7 +90,13 @@ class AnthropicStreamingInterface:
 
     def get_tool_call_object(self) -> ToolCall:
         """Useful for agent loop"""
-        return ToolCall(id=self.tool_call_id, function=FunctionCall(arguments=self.accumulated_tool_call_args, name=self.tool_call_name))
+        # hack for tool rules
+        tool_input = json.loads(self.accumulated_tool_call_args)
+        if "id" in tool_input and tool_input["id"].startswith("toolu_") and "function" in tool_input:
+            arguments = str(json.dumps(tool_input["function"]["arguments"], indent=2))
+        else:
+            arguments = self.accumulated_tool_call_args
+        return ToolCall(id=self.tool_call_id, function=FunctionCall(arguments=arguments, name=self.tool_call_name))
 
     def _check_inner_thoughts_complete(self, combined_args: str) -> bool:
         """
@@ -311,6 +319,7 @@ class AnthropicStreamingInterface:
                         self.message_id = event.message.id
                         self.input_tokens += event.message.usage.input_tokens
                         self.output_tokens += event.message.usage.output_tokens
+                        self.model = event.message.model
                     elif isinstance(event, BetaRawMessageDeltaEvent):
                         self.output_tokens += event.usage.output_tokens
                     elif isinstance(event, BetaRawMessageStopEvent):

letta/interfaces/openai_streaming_interface.py

@@ -40,6 +40,9 @@ class OpenAIStreamingInterface:
         self.letta_assistant_message_id = Message.generate_id()
         self.letta_tool_message_id = Message.generate_id()
 
+        self.message_id = None
+        self.model = None
+
         # token counters
         self.input_tokens = 0
         self.output_tokens = 0
@@ -69,10 +72,14 @@ class OpenAIStreamingInterface:
             prev_message_type = None
             message_index = 0
             async for chunk in stream:
+                if not self.model or not self.message_id:
+                    self.model = chunk.model
+                    self.message_id = chunk.id
+
                 # track usage
                 if chunk.usage:
-                    self.input_tokens += len(chunk.usage.prompt_tokens)
-                    self.output_tokens += len(chunk.usage.completion_tokens)
+                    self.input_tokens += chunk.usage.prompt_tokens
+                    self.output_tokens += chunk.usage.completion_tokens
 
                 if chunk.choices:
                     choice = chunk.choices[0]

letta/llm_api/anthropic.py

@@ -134,13 +134,13 @@ def anthropic_check_valid_api_key(api_key: Union[str, None]) -> None:
 
 
 def antropic_get_model_context_window(url: str, api_key: Union[str, None], model: str) -> int:
-    for model_dict in anthropic_get_model_list(url=url, api_key=api_key):
+    for model_dict in anthropic_get_model_list(api_key=api_key):
         if model_dict["name"] == model:
             return model_dict["context_window"]
     raise ValueError(f"Can't find model '{model}' in Anthropic model list")
 
 
-def anthropic_get_model_list(url: str, api_key: Union[str, None]) -> dict:
+def anthropic_get_model_list(api_key: Optional[str]) -> dict:
     """https://docs.anthropic.com/claude/docs/models-overview"""
 
     # NOTE: currently there is no GET /models, so we need to hardcode
@@ -159,6 +159,25 @@ def anthropic_get_model_list(url: str, api_key: Union[str, None]) -> dict:
     return models_json["data"]
 
 
+async def anthropic_get_model_list_async(api_key: Optional[str]) -> dict:
+    """https://docs.anthropic.com/claude/docs/models-overview"""
+
+    # NOTE: currently there is no GET /models, so we need to hardcode
+    # return MODEL_LIST
+
+    if api_key:
+        anthropic_client = anthropic.AsyncAnthropic(api_key=api_key)
+    elif model_settings.anthropic_api_key:
+        anthropic_client = anthropic.AsyncAnthropic()
+    else:
+        raise ValueError("No API key provided")
+
+    models = await anthropic_client.models.list()
+    models_json = models.model_dump()
+    assert "data" in models_json, f"Anthropic model query response missing 'data' field: {models_json}"
+    return models_json["data"]
+
+
 def convert_tools_to_anthropic_format(tools: List[Tool]) -> List[dict]:
     """See: https://docs.anthropic.com/claude/docs/tool-use
 

letta/llm_api/anthropic_client.py

@@ -35,6 +35,7 @@ from letta.schemas.openai.chat_completion_response import ChatCompletionResponse
 from letta.schemas.openai.chat_completion_response import Message as ChoiceMessage
 from letta.schemas.openai.chat_completion_response import ToolCall, UsageStatistics
 from letta.services.provider_manager import ProviderManager
+from letta.settings import model_settings
 from letta.tracing import trace_method
 
 DUMMY_FIRST_USER_MESSAGE = "User initializing bootup sequence."
@@ -120,8 +121,16 @@ class AnthropicClient(LLMClientBase):
             override_key = ProviderManager().get_override_key(llm_config.provider_name, actor=self.actor)
 
         if async_client:
-            return anthropic.AsyncAnthropic(api_key=override_key) if override_key else anthropic.AsyncAnthropic()
-        return anthropic.Anthropic(api_key=override_key) if override_key else anthropic.Anthropic()
+            return (
+                anthropic.AsyncAnthropic(api_key=override_key, max_retries=model_settings.anthropic_max_retries)
+                if override_key
+                else anthropic.AsyncAnthropic(max_retries=model_settings.anthropic_max_retries)
+            )
+        return (
+            anthropic.Anthropic(api_key=override_key, max_retries=model_settings.anthropic_max_retries)
+            if override_key
+            else anthropic.Anthropic(max_retries=model_settings.anthropic_max_retries)
+        )
 
     @trace_method
     def build_request_data(
@@ -239,6 +248,24 @@ class AnthropicClient(LLMClientBase):
 
         return data
 
+    async def count_tokens(self, messages: List[dict] = None, model: str = None, tools: List[Tool] = None) -> int:
+        client = anthropic.AsyncAnthropic()
+        if messages and len(messages) == 0:
+            messages = None
+        if tools and len(tools) > 0:
+            anthropic_tools = convert_tools_to_anthropic_format(tools)
+        else:
+            anthropic_tools = None
+        result = await client.beta.messages.count_tokens(
+            model=model or "claude-3-7-sonnet-20250219",
+            messages=messages or [{"role": "user", "content": "hi"}],
+            tools=anthropic_tools or [],
+        )
+        token_count = result.input_tokens
+        if messages is None:
+            token_count -= 8
+        return token_count
+
     def handle_llm_error(self, e: Exception) -> Exception:
         if isinstance(e, anthropic.APIConnectionError):
             logger.warning(f"[Anthropic] API connection error: {e.__cause__}")
@@ -369,11 +396,11 @@ class AnthropicClient(LLMClientBase):
                     content = strip_xml_tags(string=content_part.text, tag="thinking")
                 if content_part.type == "tool_use":
                     # hack for tool rules
-                    input = json.loads(json.dumps(content_part.input))
-                    if "id" in input and input["id"].startswith("toolu_") and "function" in input:
-                        arguments = str(input["function"]["arguments"])
+                    tool_input = json.loads(json.dumps(content_part.input))
+                    if "id" in tool_input and tool_input["id"].startswith("toolu_") and "function" in tool_input:
+                        arguments = str(tool_input["function"]["arguments"])
                     else:
-                        arguments = json.dumps(content_part.input, indent=2)
+                        arguments = json.dumps(tool_input, indent=2)
                     tool_calls = [
                         ToolCall(
                             id=content_part.id,