Enable Streaming on an LLM
Wire stream=True on the
LLM
class so token chunks emit while the completion is still in flight.
from crewai import LLM, Agent
llm = LLM(model="openai/gpt-4o", stream=True)
agent = Agent(role="...", goal="...", backstory="...", llm=llm)Chunk Types
- •
LLMStreamChunkEvent— partial token chunks. - •
LLMThinkingChunkEvent— extended-thinking reasoning chunks (when supported).
Forward Chunks to a UI
Subscribe a BaseEventListener and push to your transport (SSE, WebSocket).
class StreamRelay(BaseEventListener):
def on_event(self, event):
if event.__class__.__name__ == "LLMStreamChunkEvent":
sse_send(event.chunk)Streaming with Flows
Construct
Flow
with stream=True. Each method's output streams to listeners. Combine with @persist for resumable streaming sessions.
Streaming with kickoff_for_each
Async batches stream results per-item as they complete. Use
kickoff_for_each()
or
kickoff_for_each_async()
from the crew API. Tune max_rpm on the Crew to avoid 429s.
Back-pressure
UI listeners that block the event thread will stall the run. Push chunks onto a queue and drain off-thread.
Kickoff entry points for streaming crews
Listeners behave the same whether you start work with kickoff() or kickoff_async(). Async services usually pair kickoff_async with tools that implement BaseTool._arun() so httpx-style I/O does not block the thread that is draining your chunk queue.
CrewAI streaming FAQ
How do I enable streaming on a CrewAI LLM?
Construct the LLM with stream=True so token chunks emit as events your listeners can forward to SSE, WebSockets, or logs without waiting for the full completion string.
What CrewAI events carry streamed LLM tokens?
LLMStreamChunkEvent surfaces partial tokens while LLMThinkingChunkEvent covers extended-thinking style chunks when the provider supports them.
How should I forward CrewAI stream chunks to a web UI?
Subscribe with BaseEventListener, push chunks onto a thread-safe queue, and drain that queue on a worker thread so slow clients cannot block the crew event loop.
Does CrewAI Flow support streaming?
Yes. Instantiate Flow with stream=True so each step can stream intermediate output while listeners record progress for dashboards or resumable sessions with @persist.
What is back-pressure when streaming CrewAI output?
Back-pressure means your consumer must keep up with producers. Blocking inside on_event stalls the run, so decouple network IO from the listener thread via queues or async tasks.
Where can I read more CrewAI streaming APIs?
Use the DevShelfHub CrewAI API reference for BaseEventListener and related classes, then revisit this tutorial while wiring observability and lite agent patterns.