LangChain

LangChain and LangGraph work with TokenPolice after one install step. Init as usual — see init parameters — there is nothing LangChain-specific to call. The provider calls underneath (OpenAI, Anthropic, Gemini, …) are metered through LangChain, and tool calls get their own steps in the trace automatically.

Python

Install the langchain extra:

pip install "token-police[langchain]"

Then tp.init(...) before you import the module that imports LangChain:

import token_police as tp

tp.init(api_key=..., base_url=..., firewall="enforce")

import agent   # imports LangChain — must come after init

Node

Add the companion package alongside the SDK. There is no API to call and nothing to import from it — the base SDK detects it at startup and enables LangChain instrumentation:

npm i token-police token-police-langchain

On native ESM you also need two static imports before init(), so the SDK can patch the ESM copies of @langchain/core that LangGraph actually calls:

import * as tp from "token-police";
import * as lcChatModels from "@langchain/core/language_models/chat_models";
import * as lcCallbacks from "@langchain/core/callbacks/manager";

tp.init({
  /* … see init parameters … */
  instrumentModules: {
    langChain: { chatModelsModule: lcChatModels, callbackManagerModule: lcCallbacks },
  },
});

// Static ESM imports all run before this file's code, so keep every other
// @langchain/* import out of the entry module and load it after init instead.
const { runTurn } = await import("./agent.js");

Supported: @langchain/core 0.3.x and 1.x (tested with langchain 0.3 and @langchain/langgraph 1.x). On @langchain/core 0.3, pnpm reports an unmet peer from the companion's bundled instrumentor (it wants >=1.0.0); this is expected and safe — the SDK wires LangChain's callback manager itself.

Next