Recipe

You have: a LangGraph application. You want: to run a known, deterministic multi-tool path inside one node — without rewriting your graph.

Paired script: examples/integrations/langgraph_node.py.

ChainWeaver is not an alternative to LangGraph. LangGraph owns open-ended, model-driven graph control; ChainWeaver owns the deterministic internal tool paths. When a node needs to run such a path, it calls executor.execute_flow(...) and merges the typed result back into graph state.

Install

LangGraph is an optional extra:

pip install 'chainweaver[langgraph]'

The boundary

flowchart LR
  START --> N[LangGraph node] --> END
  N -.calls.-> FW[[ChainWeaver flow: normalize -> word_count]]

The node

def run_flow_node(state: GraphState) -> dict[str, Any]:
    result = executor.execute_flow("enrich_text", {"text": state["raw_text"]})
    if not result.success or result.final_output is None:
        raise RuntimeError("enrich_text flow failed")
    return {
        "normalized": result.final_output["normalized"],
        "word_count": result.final_output["word_count"],
    }

graph = StateGraph(GraphState)
graph.add_node("run_flow", run_flow_node)
graph.add_edge(START, "run_flow")
graph.add_edge("run_flow", END)
app = graph.compile()

The node passes selected state fields into the flow and merges selected flow outputs back out. LangGraph decides whether to enter the node; ChainWeaver runs the path deterministically once inside.

When this pattern fits

• Good fit: a sub-path of your graph is a fixed sequence of tool calls with no model decision in between.
• Not a fit: the next step depends on a model's judgement — keep that in LangGraph; only the deterministic stretch belongs in a flow.

What next

• OpenAI Agents SDK recipe — the same idea for a different runtime.
• The broader ecosystem bridge adapters are tracked in issue #82.