docs/edge/en/concepts/streaming.mdx
Streaming lets your application receive execution updates while work is still running. Instead of waiting for the final result, you can render LLM tokens, tool activity, Flow lifecycle events, and conversation messages as they happen.
CrewAI has two streaming surfaces:
| Surface | Used by | Output |
|---|---|---|
| Frame streaming | Flows, direct LLM calls, conversational turns | Ordered StreamFrame objects |
| Crew chunk streaming | Crews with stream=True | CrewStreamingOutput chunks |
For new runtime integrations, UIs, terminal apps, service bridges, and conversational surfaces, use frame streaming. It provides one stable event envelope across the runtime.
A StreamFrame is the common object emitted by streamable runtimes:
frame.id # unique frame id
frame.seq # execution-local order, when available
frame.type # source event type, such as "llm_stream_chunk"
frame.channel # "llm", "flow", "tools", "messages", "lifecycle", or "custom"
frame.namespace # source/runtime namespace
frame.timestamp # event timestamp
frame.parent_id # parent event id, when available
frame.previous_id # previous event id, when available
frame.data # structured event payload
frame.event # alias for frame.data
frame.content # printable text for token-like frames, otherwise ""
The important fields for most consumers are:
| Field | Use it for |
|---|---|
channel | Routing frames to the right UI region |
type | Handling a specific event inside a channel |
content | Printing token-like text |
event | Reading structured metadata, such as tool names or message roles |
seq | Preserving execution order |
Frames are grouped into high-level channels:
| Channel | Contains |
|---|---|
llm | LLM call lifecycle, text chunks, and thinking chunks |
flow | Flow lifecycle, method execution, routing, pause, and resume events |
tools | Tool usage start, finish, and error events |
messages | Conversation transcript events |
lifecycle | Runtime lifecycle events that do not belong to another channel |
custom | Events that do not map to a built-in channel |
The stream itself remains one ordered timeline. Channel projections let consumers focus on only part of that timeline.
flowchart LR
A["flow
flow_started"] --> B["llm
llm_call_started"]
B --> C["llm
llm_stream_chunk"]
C --> D["tools
tool_usage_started"]
D --> E["tools
tool_usage_finished"]
E --> F["llm
llm_stream_chunk"]
F --> G["flow
flow_finished"]
Frame streaming returns a stream session:
stream = flow.stream_events(inputs={"topic": "AI agents"})
The session is both an iterator and the holder for the final result:
with stream:
for frame in stream:
print(frame.content, end="", flush=True)
result = stream.result
Consume the stream before reading stream.result. Reading the result too early raises an error because the runtime may still be producing frames.
Use channel projections when you only need one kind of frame:
with flow.stream_events(inputs={"topic": "AI agents"}) as stream:
for frame in stream.llm:
print(frame.content, end="", flush=True)
result = stream.result
Available projections:
| Projection | Frames |
|---|---|
stream.events | All frames |
stream.llm | LLM frames |
stream.flow | Flow frames |
stream.tools | Tool frames |
stream.messages | Conversation message frames |
stream.interleave([...]) | Selected channels in relative order |
Use the entrypoint that matches the runtime you are streaming:
| Runtime | Streaming entrypoint |
|---|---|
| Flow | flow.stream_events(...) |
Flow with stream=True | flow.kickoff(...) returns a stream session |
| Async Flow | flow.astream(...) or await flow.kickoff_async(...) when stream=True |
| Direct LLM call | llm.stream_events(...) |
| Conversational Flow turn | flow.stream_turn(...) |
| Crew | Crew(..., stream=True).kickoff(...) returns CrewStreamingOutput |
Direct llm.call(...) still returns the final assembled LLM result. Use llm.stream_events(...) when you want to iterate over LLM chunks as they arrive.