LangGraph Adapter

[zahlekhan]

Is your feature request related to a problem? Please describe.
OpenUI currently provides streaming adapters for OpenAI (Chat Completions, Responses API, Readable Stream) and AG-UI, but there is no adapter for LangGraph. Users building agentic applications with LangGraph cannot easily connect their LangGraph backend to OpenUI's streaming renderer and chat interfaces.

Describe the solution you'd like
Add a new langGraphAdapter in @openuidev/react-headless that implements the StreamProtocolAdapter interface for LangGraph's SSE-based streaming protocol. The adapter should:

• Parse LangGraph Cloud / LangGraph SDK streaming responses (SSE events like metadata, data, error, end).
• Support multi-agent runs — correctly handle events from multiple agents/nodes in a LangGraph graph, mapping them to the appropriate OpenUI rendering flow.
• Support LangGraph-specific features like interrupts and human-in-the-loop patterns.
• Be usable via streamProtocol="langgraph" on ChatProvider.
• Include a corresponding langGraphMessageFormat converter if the message format differs from existing ones.

Describe alternatives you've considered

• Users could write a server-side translation layer that converts LangGraph streams into a format an existing adapter (e.g., AG-UI) understands, but this adds latency and complexity.
• A generic SSE adapter could partially work, but would miss LangGraph-specific semantics like multi-agent routing and interrupts.

Additional context
Existing adapters live in packages/react-headless/src/stream/adapters/. The adapter should follow the same StreamProtocolAdapter pattern (implementing async *parse(response): AsyncIterable<AGUIEvent>). LangGraph's streaming protocol documentation: https://langchain-ai.github.io/langgraph/cloud/concepts/streaming/

[PatrickPrakash]

Is anyone actively working on it? May I can try to implement it

[alkitrend]

Hey, I've dug into the codebase and would love to implement this.

I've looked at the existing adapters (ag-ui.ts, openai-completions.ts, openai-responses.ts) and the StreamProtocolAdapter interface. The LangGraph adapter fits naturally into the same pattern, just the main complexity is in correctly handling LangGraph's named SSE events (event: messages/partial, messages/complete, updates, error, end) vs. the standard data:-only SSE the other adapters handle, plus the [message_chunk, metadata] tuple format LangGraph uses for its messages stream mode.

Here's my plan for the full event mapping:

LangGraph event	AG-UI event
metadata	RUN_STARTED
messages/partial (text delta)	TEXT_MESSAGE_START → TEXT_MESSAGE_CONTENT
messages/partial (tool_calls chunk)	TOOL_CALL_START → TOOL_CALL_ARGS
messages/complete	TEXT_MESSAGE_END / TOOL_CALL_END
updates with interrupt	RUN_FINISHED + interrupt callback
error	RUN_ERROR
end	RUN_FINISHED

For multi-agent support, the langgraph_node field in the metadata tuple maps cleanly to STEP_STARTED/STEP_FINISHED AG-UI events, which is a natural fit.

For interrupts, I'd expose an onInterrupt callback on the adapter options rather than a custom event type, keeping it consistent with how the existing adapters handle edge cases without leaking LangGraph-specific concepts into the AG-UI layer.

[zahlekhan]

@alkitrend plan looks good. I'm assigning it you.

[alkitrend]

@alkitrend plan looks good. I'm assigning it you.

I will do my best.

[alkitrend]

I implemented, please check this PR: #381