SDKs

Quick

You do not change SDKs. Install the same official package you always would; two constructor arguments change — the base URL and the key.

# The OpenAI-compatible dialect
pip install openai
npm install openai

# The native Anthropic dialect
pip install anthropic
npm install @anthropic-ai/sdk

Keep the key in an environment variable. None of the examples below writes a secret into source — they all read KUMO_API_KEY, so export it first in the shell you run the code from.

export KUMO_API_KEY="kumo_sk_..."

A new organization's wallet is empty: a first call without a top-up answers 402 (billing).

The OpenAI SDK

The base URL is https://api.kumorouter.com/v1, including the /v1. The SDK appends the operation path itself.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.kumorouter.com/v1",
    api_key=os.environ["KUMO_API_KEY"],
)

answer = client.chat.completions.create(
    model="<model>",
    max_tokens=128,
    messages=[{"role": "user", "content": "Explain tokens in one line."}],
)
print(answer.choices[0].message.content)

The output ceiling is required: a request without max_tokens (equivalently max_completion_tokens) is refused rather than left unbounded.

The Anthropic SDK

The base URL is https://api.kumorouter.com, the bare origin, with no `/v1`: the SDK appends /v1/messages itself. The key travels in the x-api-key header, and the SDK puts it there without your help.

import os
from anthropic import Anthropic

client = Anthropic(
    base_url="https://api.kumorouter.com",
    api_key=os.environ["KUMO_API_KEY"],
)

answer = client.messages.create(
    model="<model>",
    max_tokens=128,
    messages=[{"role": "user", "content": "Explain tokens in one line."}],
)
print(answer.content[0].text)

How the key header is written → What a model can do →

Streaming

Both SDKs stream, and the gateway answers each in the shape of its own protocol: events in arrival order, then a terminating frame. There is one difference from a unary call — a failure after the stream has begun cuts it off with no terminating frame, and such an answer counts as a failed call, not a short one.

Streaming in full →

Libraries built on the OpenAI client

Frameworks such as LangChain or LlamaIndex usually reach models through that same OpenAI client. If a library lets you configure the OpenAI client's base URL, it works with the gateway exactly as a direct call does: set https://api.kumorouter.com/v1 and the key, and leave everything else alone.

The exact parameter name differs per library, and it is deliberately not named here: names are renamed between releases, and a wrong name looks like a setting that was silently ignored. Check your library's own documentation — what you are looking for is the place where the OpenAI client's base URL is set.

The easiest proof that the setting took effect is the console log: the call is there, so the library went through the gateway.