LangChain - Orbitrage

LangChain’s ChatOpenAI is OpenAI-compatible — point it at the Orbitrage gateway and every call is routed and traced. Tools, managed tools, and streaming all work unchanged.

Install

pip install -U orbitrage langchain langchain-openai

npm install orbitrage@latest @langchain/openai @langchain/core

Basic chat

Wrapper clients build their own HTTP client, so set the base URL and key explicitly, and pass your end-user id via default_headers.

import os, orbitrage
orbitrage.init(os.environ["ORBITRAGE_API_KEY"], user_id="customer_42")

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="glm-5.2",                       # direct model — or "minimax-m3"; "auto" to route
    base_url="https://api.orbitrage.ai/v1",
    api_key=os.environ["ORBITRAGE_API_KEY"],
    default_headers={"x-orbitrage-end-user-id": "customer_42"},
    max_tokens=512,
)
print(llm.invoke("Explain routing in one line.").content)

import { orbitrage } from "orbitrage";
await orbitrage.init({ apiKey: process.env.ORBITRAGE_API_KEY, userId: "customer_42" });

import { ChatOpenAI } from "@langchain/openai";

const llm = new ChatOpenAI({
  model: "glm-5.2",                        // direct model — or "minimax-m3"; "auto" to route
  apiKey: process.env.ORBITRAGE_API_KEY,
  maxTokens: 512,
  configuration: {
    baseURL: "https://api.orbitrage.ai/v1",
    defaultHeaders: { "x-orbitrage-end-user-id": "customer_42" },
  },
});
console.log((await llm.invoke("Explain routing in one line.")).content);

Managed tools (web search, scrape, calculator — no keys)

Reference a managed tool by name and Orbitrage runs it server-side, then loops the result back to the model.

# bind passes the managed tool straight through to the API
out = llm.bind(tools=["calculator_orbitrage"]).invoke(
    "Use the calculator to compute 1234*5678. Only the number."
)
print(out.content)   # 7006652

// modelKwargs forwards the managed tool to the API
const withTool = new ChatOpenAI({
  model: "minimax-m3",
  apiKey: process.env.ORBITRAGE_API_KEY,
  modelKwargs: { tools: ["calculator_orbitrage"] },
  configuration: { baseURL: "https://api.orbitrage.ai/v1" },
});
const out = await withTool.invoke("Use the calculator to compute 1234*5678. Only the number.");
console.log(out.content);   // 7006652

Your own LangChain tools (@tool / bind_tools) keep working alongside managed tools in the same call.

Streaming

for chunk in llm.stream("Count from 1 to 5."):
    print(chunk.content, end="", flush=True)

for await (const chunk of await llm.stream("Count from 1 to 5.")) {
  process.stdout.write(String(chunk.content));
}

Pin a direct model (glm-5.2, minimax-m3) for predictable behavior. Use model="auto" to let Orbitrage pick the cheapest capable model — and keep max_tokens ≥ 512 so reasoning models aren’t truncated.