If you already use the official OpenAI client, switching to Corvex is a
three-line constructor change: your API key, the base URL, and the closing
of the constructor block. Everything else — request shape, response shape,
streaming, tools, models — is identical.
The examples below are the canonical sources validated in CI on every
docs PR. They are copy-paste runnable: set OPENAI_API_KEY to your virtual
key (sk-corvex-...) and the script makes one real /v1/chat/completions call
against https://inference.corvex.studio.
The 3-line swap
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["OPENAI_API_KEY"],
base_url="https://inference.corvex.studio/v1",
)
print(client.chat.completions.create(
model="Qwen/Qwen2.5-0.5B-Instruct",
messages=[{"role": "user", "content": "Hello, Corvex."}],
).choices[0].message.content)
Authorization: Bearer,
just like OpenAI.
Run it
# Python
pip install openai
export OPENAI_API_KEY=sk-corvex-your-virtual-key
python python.py
# Node (requires Node 18+ for native fetch + top-level await)
npm install openai
export OPENAI_API_KEY=sk-corvex-your-virtual-key
node node.mjs
APIError (Python) or OpenAI.APIError (Node)
with the gateway’s structured error body.
Model naming
Corvex uses bare lab-canonical Hugging Face repo ids for every model
(e.g. zai-org/GLM-5.1-FP8, deepseek-ai/DeepSeek-V4-Flash). There is no Corvex
prefix; legacy corvex/<…> and corvex-pd/<…> ids return HTTP 404 since
the Day-1 dogfooding cutover (ADR 024 / RD-651). PD vs regular topology
is decided server-side from the catalog, not from the request. List the
models your virtual key can access:
curl https://inference.corvex.studio/v1/models \
-H "Authorization: Bearer sk-corvex-YOUR_VIRTUAL_KEY"
What’s next
- Quickstart → — broader guide
including the curl reference and authentication notes.
- Authentication → — virtual
keys vs admin keys; which to use for what.
- API Reference → — every endpoint,
request/response schemas, and error codes.
