LangChain and LangGraph

Track token usage and estimated model costs in LangChain and LangGraph, attribute calls to customers and features, and check Veqen budgets before execution.

The langchain-veqen package exposes Veqen's maintained callback handler. Attach it to a model or graph invocation to record usage without prompt or response text.

pip install langchain-veqen langchain-openai

import veqen
from langchain_openai import ChatOpenAI
from langchain_veqen import VeqenCallbackHandler

veqen.init(ledger_file="ai-costs.jsonl")
model = ChatOpenAI(model="gpt-4o-mini", stream_usage=True)
answer = model.invoke(
    "Explain the difference between input and output tokens.",
    config={"callbacks": [VeqenCallbackHandler(customer="demo", feature="assistant")]},
)
veqen.flush()

Set OPENAI_API_KEY for this example. Local recording needs no Veqen account. To also send usage to your organization, set VEQEN_API_KEY to an ingest-only key before starting the process. Create it under Settings → API keys, following the quickstart, and keep it on the server.

LangGraph

Pass the handler through the callback configuration on your graph invocation so it reaches nested model calls:

result = graph.invoke(
    inputs,
    config={"callbacks": [VeqenCallbackHandler(customer="demo", project="research-agent")]},
)

Here graph is your compiled graph and inputs is its normal input. Use one recording path per model call: adding this handler and wrapping the same provider client can record the call twice.

Usage and budget behavior

The handler reads standard AIMessage.usage_metadata, model identifiers and provider metadata. Enable streaming usage in your model integration when streaming. Calls without usage metadata or a model identifier are skipped rather than assigned guessed counts. Cache reads and cache writes are separated from uncached input tokens. This callback handles token-based chat usage; use the Python SDK for provider-specific media billing fields.

Cost uses the versioned price book. It remains an estimate until reconciled with the provider bill. Local budgets and cached cloud budget decisions are checked before model execution; budget documentation explains refresh intervals and concurrency limits.

The same class remains available as veqen.integrations.langchain.VeqenCallbackHandler for existing veqen[langchain] installations. Source and tests are public.