LangGraph

Connect any LangGraph workflow to MeshFlow in two steps: install the adapter and wrap your compiled graph. Every node execution emits an ATR-compatible span — no graph rewrite required.

Zero rewrite

The adapter hooks into LangGraph's native callback system. Your graph definition, nodes, and edges stay exactly as-is.

Install

bash

pip install "meshflow[langgraph]"

Callback handler

Pass MeshFlowCallbackHandler to your graph's config. It automatically emits spans for every node start, end, and error.

python

from meshflow.adapters.langgraph import MeshFlowCallbackHandler

from meshflow.cloud import MeshFlowCloud

cloud = MeshFlowCloud(api_key="mf-...")

handler = MeshFlowCallbackHandler(cloud=cloud, run_id="run-abc123")

# Pass to your compiled graph — existing graph unchanged

inputs = {"messages": [HumanMessage(content="your task")]}

cfg = {"callbacks": [handler]}

result = app.invoke(inputs, config=cfg)

ATR fields

The handler auto-populates modelVersion, inputHash, outputHash, and causalParentId on every span using LangGraph's run metadata.

python

# Override ATR fields per-node if needed

handler.set_node_meta("call_model", prompt_version="v12", policy_version="hipaa-v3")

Context packs

Emit a context snapshot before each LLM call to capture which sources, memory reads, and prompt versions the node received.

python

from meshflow.adapters.langgraph import emit_context_pack

def call_model(state):

emit_context_pack(cloud=cloud, state=state, agent_name="call_model")

return {"messages": [llm.invoke(state["messages"])]}

OpenAI Agents SDK

CrewAI

LangGraph

Connect any LangGraph workflow to MeshFlow in two steps: install the adapter and wrap your compiled graph. Every node execution emits an ATR-compatible span — no graph rewrite required.

Zero rewrite

The adapter hooks into LangGraph's native callback system. Your graph definition, nodes, and edges stay exactly as-is.

Callback handler

Pass MeshFlowCallbackHandler to your graph's config. It automatically emits spans for every node start, end, and error.

python

from meshflow.adapters.langgraph import MeshFlowCallbackHandler

from meshflow.cloud import MeshFlowCloud

cloud = MeshFlowCloud(api_key="mf-...")

handler = MeshFlowCallbackHandler(cloud=cloud, run_id="run-abc123")

# Pass to your compiled graph — existing graph unchanged

inputs = {"messages": [HumanMessage(content="your task")]}

cfg = {"callbacks": [handler]}

result = app.invoke(inputs, config=cfg)

Context packs

Emit a context snapshot before each LLM call to capture which sources, memory reads, and prompt versions the node received.

python

from meshflow.adapters.langgraph import emit_context_pack

def call_model(state):

emit_context_pack(cloud=cloud, state=state, agent_name="call_model")

return {"messages": [llm.invoke(state["messages"])]}