All posts
How OpenAI-compatible endpoints actually work
Engineering 7 min read· 5 Jul 2026· By Engineering

How OpenAI-compatible endpoints actually work

Under the hood of the base_url swap: how a single URL change routes your request through any provider.

The OpenAI SDK is really just a typed HTTP client. Every call boils down to a POST to /v1/chat/completions with a JSON body. If a server speaks that shape, the SDK cannot tell the difference.

The base_url trick#

python
from openai import OpenAI

client = OpenAI(
    api_key="sk-tb-...",
    base_url="https://tokenbaazar.in/api/public/v1",
)

Every subsequent method call — chat.completions.create, embeddings.create, images.generate — hits our proxy instead of api.openai.com.

What we do at the edge#

  • Validate your sk-tb- key against a hashed record
  • Check wallet balance for the requested model
  • Route to the upstream provider (Gemini, OpenAI, Anthropic tier)
  • Deduct the exact per-token cost after the response streams
Edge routing diagram
Requests are metered at the edge, not buffered.

Why it matters#

You get zero lock-in. If you ever want to leave TokenBaazar, change the base_url back and your code still runs. Read the full migration guide.

Streaming works identically — we pass SSE frames through as they arrive, so first-token latency is upstream latency plus ~40ms.

Ready to build?

Grab a key, keep your OpenAI SDK, and pay in INR.

Start free