LangChain

Install:

shell

pip install "nullrun[langchain]"

langchain_chat.py

import nullrun
from langchain_openai import ChatOpenAI
from langchain_core.messages import HumanMessage

nullrun.init(api_key="nr_live_...")

llm = ChatOpenAI(model="gpt-4o-mini")
resp = llm.invoke([HumanMessage(content="Hello")])

nullrun.init() patches BaseCallbackManager.__init__ via patch_langchain_callback in nullrun.instrumentation.auto, so a NullRunCallback is added to every CallbackManager on construction. That callback fires the same track_llm path the httpx hook uses, so cost tracking works for in-memory mock providers and callback-only flows that don't hit the network (the httpx hook alone only covers networking calls).

The same auto-instrumentation path works for any LangChain Runnable — chains, agents, retrievers. For LangGraph specifically, see Protect a LangGraph agent — Pregel.invoke / Pregel.stream get an extra wrapper layer via patch_langgraph_compiled.

Test coverage

Per NullRun's testing policy (see the SDK README), LangChain callback patching exists but has no dedicated callback-integration test. The patch_langchain_callback hook is exercised by the unified-fingerprint test (tests/test_unified_fingerprint.py), which validates the fingerprint shape but not the callback's wire effect. This is documented per §8 of the source-of-truth positioning — verify behaviour against your real workload before relying on the callback path.