LangGraph

LangGraph builds on langchain-openai’s ChatOpenAI, so the integration shape is identical to LangChain — configure the LLM with base_url + default_headers, then use it as any node’s LLM.

Minimal example

import os
from langchain_openai import ChatOpenAI
from langgraph.graph import StateGraph, END
 
llm = ChatOpenAI(
    base_url=os.environ.get("CM_GATEWAY_URL", "https://api.curate-me.ai/v1/openai"),
    default_headers={"X-CM-API-Key": os.environ["CM_API_KEY"]},
    model="gpt-4o",
)
 
def respond(state):
    return {"messages": state["messages"] + [llm.invoke(state["messages"])]}
 
graph = StateGraph(dict)
graph.add_node("respond", respond)
graph.set_entry_point("respond")
graph.add_edge("respond", END)
app = graph.compile()

Full single-node and multi-agent examples live at examples/integrations/langgraph/.

What’s traced

Each node’s LLM call shows up as its own trace row in the dashboard, keyed by X-CM-Request-Id. The trace carries the node name as a custom attribute when you set tags=["node:respond"] on the LLM invocation — this lets you slice cost by node.

Supported versions

Known limitations

• Checkpointer: the gateway only sees per-node LLM calls. State snapshots persisted by MemorySaver / PostgresSaver / SqliteSaver run entirely on your infrastructure. We don’t yet store cross-step memory diffs.
• Conditional routing: edges that don’t make LLM calls (pure-Python routers) don’t show up in the trace — that’s by design, the gateway is invoked only when an LLM call happens.
• Subgraphs: each subgraph node is a separate trace. Use the W3C traceparent header on outer calls to link them.