Using from LiteLLM

LiteLLM talks to Oraicle as an OpenAI-compatible provider. Nothing needs a custom handler: prefix the model with openai/, set the base URL, and pass your Oraicle key.

Get a key from the Dashboard. Keys begin with oai-.

The proxy

Add the models you want to config.yaml. The model_name is what your application asks for; the model under litellm_params is Oraicle’s id, with openai/ in front so LiteLLM uses its OpenAI-compatible path.

model_list:
  - model_name: glm-4.7-flash
    litellm_params:
      model: openai/zai-org/GLM-4.7-Flash
      api_base: https://api.oraicle.me/v1
      api_key: os.environ/ORAICLE_API_KEY

  - model_name: gpt-oss-120b
    litellm_params:
      model: openai/openai/gpt-oss-120b
      api_base: https://api.oraicle.me/v1
      api_key: os.environ/ORAICLE_API_KEY

The double prefix is not a typo

openai/openai/gpt-oss-120b reads oddly and is correct. The first openai/ tells LiteLLM which handler to use; the rest is the model id exactly as Oraicle serves it — and that model’s id happens to start with openai/ too. Getting this wrong gives a 404 from us, not an error from LiteLLM.

Then:

export ORAICLE_API_KEY=oai-...
litellm --config config.yaml

Your application points at the proxy and asks for glm-4.7-flash.

The Python SDK

Without the proxy, pass the same three values to completion():

import os
from litellm import completion

response = completion(
    model="openai/zai-org/GLM-4.7-Flash",
    api_base="https://api.oraicle.me/v1",
    api_key=os.environ["ORAICLE_API_KEY"],
    messages=[{"role": "user", "content": "Hello"}],
)

print(response.choices[0].message.content)

Streaming works the same way — pass stream=True.

Which models to list

Every id from Models can go in the config. For the current list at any moment:

curl https://api.oraicle.me/v1/models

Each entry carries pricing.input and pricing.output in US dollars per million tokens, from the same table billing reads.

Do not point LiteLLM's embeddings at us yet

Only chat completions and live-data oracles are live. There is no /v1/embeddings endpoint, so an embedding entry in your config will fail when it is called. See What’s live and what’s next.

Costs and limits

Every call is billed per token from your prepaid Oraicle wallet — no subscription, and no per-model plan to choose. LiteLLM’s own max_budget and rate-limit settings still apply on top, and are worth setting if the proxy is shared.

A call that fails before producing anything — a bad model id, an upstream outage — is not billed. A reply that starts streaming and then breaks is billed for the tokens it did produce, because the model produced them. A wallet with no balance returns an error rather than queueing.

Live data from LiteLLM

Oraicle’s oracles — web, news, weather, stocks, crypto, maps, images — are switched on with a tools entry of the shape {"type": "weather"}. That is an extension to the OpenAI request, so LiteLLM will not produce it for you.

To use one, pass it through as an extra body parameter:

response = completion(
    model="openai/openai/gpt-oss-120b",
    api_base="https://api.oraicle.me/v1",
    api_key=os.environ["ORAICLE_API_KEY"],
    messages=[{"role": "user", "content": "What's the weather in Tokyo?"}],
    extra_body={"tools": [{"type": "weather"}]},
)

Standard OpenAI function tools keep working alongside these, and LiteLLM handles those the usual way. Each oracle lookup is billed once on top of the tokens, and only when it succeeds — see Live data (oracles).