Disabling state streaming

Granularly control what is streamed to the frontend.

What is this?#

By default, CopilotKit will stream both your state and tool calls to the frontend. You can disable this by using CopilotKit's custom RunnableConfig.

When should I use this?#

Occasionally, you'll want to disable streaming temporarily — for example, the LLM may be doing something the current user should not see, like emitting tool calls or questions pertaining to other employees in an HR system.

Implementation#

Disable all streaming#

You can disable all message streaming and tool call streaming by passing emit_messages=False and emit_tool_calls=False to the CopilotKit config.

from copilotkit.langgraph import copilotkit_customize_config

async def frontend_actions_node(state: AgentState, config: RunnableConfig):

    # 1) Configure CopilotKit not to emit messages
    modifiedConfig = copilotkit_customize_config(
        config,
        emit_messages=False, # if you want to disable message streaming # [!code highlight]
        emit_tool_calls=False # if you want to disable tool call streaming # [!code highlight]
    )

    # 2) Provide the actions to the LLM
    model = ChatOpenAI(model="gpt-5.4").bind_tools([
      *state["copilotkit"]["actions"],
      # ... any tools you want to make available to the model
    ])

    # 3) Call the model with CopilotKit's modified config  # [!code highlight]
    response = await model.ainvoke(state["messages"], modifiedConfig) # [!code highlight]

    # don't return the new response to hide it from the user
    return state

BEWARE!

In LangGraph Python, the config variable in the surrounding namespace is implicitly passed into LangChain LLM calls, even when not explicitly provided.

This is why we create a new variable modifiedConfig rather than modifying config directly. If we modified config itself, it would change the default configuration for all subsequent LLM calls in that namespace.

# if we override the config variable name with a new value
config = copilotkit_customize_config(config, ...)

# it will affect every subsequent LangChain LLM call in the same namespace, even when `config` is not explicitly provided
response = await model2.ainvoke(*state["messages"]) # implicitly uses the modified config!