create_react_agent.stream not working correctly with MCP tools

[nomiku1999]

async def stream(self, query, sessionId) -> AsyncIterable[dict[str, Any]]:
    async with MultiServerMCPClient(
        {
            "ibcf": {
                "transport": "sse",
                "url": "http://localhost:7000/sse"
            },
        }
    ) as client:
        graph = create_react_agent(
            self.model,
            tools=client.get_tools(),
            checkpointer=memory,
            prompt=self.SYSTEM_INSTRUCTION,
            response_format=ResponseFormat,

        )

        inputs = {'messages': [('user', query)]}
        config = {'configurable': {'thread_id': sessionId}}

        for item in graph.stream(inputs, config, stream_mode='values'):
            logger.debug("Looping through the stream")
            
            message = item['messages'][-1]
            logger.debug(message.pretty_repr())
            if (
                isinstance(message, AIMessage)
                and message.tool_calls
                and len(message.tool_calls) > 0
            ):
                yield {
                    'is_task_complete': False,
                    'require_user_input': False,
                    'content': 'Looking up the IBCF response...',
                }
            elif isinstance(message, ToolMessage):
                yield {
                    'is_task_complete': False,
                    'require_user_input': False,
                    'content': 'Processing the IBCF response..',
                }

The tools execution keep return the error:
"""Error: NotImplementedError('StructuredTool does not support sync invocation.')
Please fix your mistakes."""

Is there anyway to fix it.
Thanks

[vbarda]

you need to use .astream() - sync invocation of tools is not supported

e.g., async for item in graph.astream

[shadowtechn]

@vbarda Hi, Just to clarify - does this mean that every time a React agent uses tools from MCP, it must be executed within an async context? If so, that would require making the entire call stack async.

Is this the recommended best practice?
Thanks

[nikil1506]

you need to use .astream() - sync invocation of tools is not supported

e.g., async for item in graph.astream

What if I use graph.invoke instead of stream?
I got the same error when I used invoke. Do I always have to use astream?

[sawradip]

You can use .ainvoke(...) mate.

https://langchain-ai.github.io/langgraph/agents/run_agents/

[nikil1506]

You can use .ainvoke(...) mate.

https://langchain-ai.github.io/langgraph/agents/run_agents/

No I figured it out. My issue was because Im a dumbass who imported create_react_agent from both the langchain and langgraph package and my graph took the one from the langchain package.

[Ch1ldKing]

Well, does anyone find that, in fact, StructruedTool supports sync invocation. Here is an example from langgraph official ly:
好吧，但是否有人发现，实际上 StructruedTool 支持 stream 和 invoke。这有一个来自 langgraph 官方的例子:

example 例子

You can find in here: https://github.com/langchain-ai/langgraph-bigtool
The code is in the Quickstart of this bigtool repository:

all_tools = []
for function_name in dir(math):
    ):
        all_tools.append(tool)
print(all_tools) # Here I tried output, the tools are  in type of StucturedTool. 我尝试了输出，这些工具的类型是StucturedTool
...
llm = init_chat_model("openai:gpt-4o-mini")

builder = create_agent(llm, tool_registry)
agent = builder.compile(store=store)
query = "Use available tools to calculate arc cosine of 0.5."

for step in agent.stream(  # Here use stream. 这里用了 stream
    {"messages": query},
    stream_mode="updates",
):
    for _, update in step.items():
        for message in update.get("messages", []):
            message.pretty_print()

It can run normally, with the response from tool.
它能正常运行，并且能收到工具调用的消息

More Infromation 更多相关信息

And in the package code langgraph_bigtool.graph.py -> create_agent, I check it how to call a tool:
并且，我查看了 Bigtool 的包代码（一个 graph.py 文件，其中的 create_agent方法），我检查了他怎么调用tool：

    def select_tools(
        tool_calls: list[dict], config: RunnableConfig, *, store: BaseStore
    ) -> State:
        selected_tools = {}
        for tool_call in tool_calls:
            kwargs = {**tool_call["args"]}
            if store_arg:
                kwargs[store_arg] = store
            result = retrieve_tools.invoke(kwargs) #Here use invoke 这里用了invoke来调用工具
            selected_tools[tool_call["id"]] = result

        tool_messages, tool_ids = _format_selected_tools(selected_tools, tool_registry)
        return {"messages": tool_messages, "selected_tool_ids": tool_ids}

Does someone can help connect the developer to solve this question? Or if you have the same problem, welcome talk about it.
有人可以帮助联系开发者来解决这个问题吗，或者有同样的问题，我们一起讨论