Skip to main content
init() points the OpenAI client at the gateway for you — use the SDK exactly as you do today. 
import os, orbitrage
orbitrage.init(os.environ["ORBITRAGE_API_KEY"], user_id="customer_42")

from openai import OpenAI
client = OpenAI()                       # base_url + auth handled by Orbitrage
client.chat.completions.create(
    model="grok-4-fast",                       # or pin: "gpt-5.4", "claude-sonnet-4-6"
    messages=[{"role": "user", "content": "hi"}],
)

​
How a bare OpenAI() reaches Orbitrage

You don’t pass a base_url — init() does it for you. On init() the SDK patches the OpenAI / AsyncOpenAI constructor so that, when you don’t supply your own base_url, every client you build:
  1. gets base_url = https://api.orbitrage.ai/v1, and
  2. is authenticated with your orb_ key — even if OPENAI_API_KEY is set in your environment. (The gateway only accepts orb_ keys; your provider keys never leave your machine.)
So a plain OpenAI() always comes to Orbitrage with the right key: 
import orbitrage; orbitrage.init("orb_...")           # OPENAI_API_KEY can be set — it's ignored for routing
from openai import OpenAI
c = OpenAI()
assert str(c.base_url).startswith("https://api.orbitrage.ai")   # ✅ routed to us
assert str(c.api_key).startswith("orb_")                        # ✅ orb key used, not sk-...
If you pass your own base_url, Orbitrage never overrides it — you stay in control.

​
Managed tools

Reference a managed tool by name and Orbitrage runs it server-side — no keys, no tool loop: 
resp = client.chat.completions.create(
    model="grok-4-fast",
    messages=[{"role": "user", "content": "Compute 1234*5678. Only the number."}],
    tools=["calculator_orbitrage"],
)
print(resp.choices[0].message.content)   # 7006652
  • model: "auto" routes to the cheapest capable model. Any model id pins it.
  • Streaming, tools, and vision all work unchanged — see Tool calling and Streaming.