Beta docs: the shipped API is POST /api/v1/generate for gemini and sarvam. Some pages below describe planned features. See the shipped quick start.

Introduction

Architecture

What sits between your application and the models it calls — and why.

OpenLink is a proxy in the classic sense: your request goes in unchanged, an answer comes back in one shape. The gateway is where provider SDKs, auth, retries, and billing stop being your problem.

Where OpenLink sits

topology
your application
      │   HTTPS /v1/chat   (one key)
      ▼
┌───────────────────────────────┐
│   OPENLINK GATEWAY · in-1     │
│   route → normalize → retry   │
└──────────┬─────────┬──────────┘
           │         │
      ┌────┴──┐  ┌───┴────┐ ┌──────────┐
      │claude │  │ gemini │ │ sarvam   │
      │ antrp │  │ google │ │ in-1     │
      └───────┘  └────────┘ └──────────┘

The gateway talks provider SDKs on the south side so your application never has to. It also terminates and re-issues every request, which is what makes per-request routing possible.

Routing

When a request arrives without a forced provider, the gateway weighs four signals — healthy availability, latency, cost, and whether a provider can answer in the message's language. The decision is made per request, and the winner is reported back in the provider field.

terminal
$ curl https://openlink-1.onrender.com/v1/chat \
    -H "Authorization: Bearer $OPENLINK_API_KEY" \
    -d '{"messages":[{"role":"user","content":"नमस्ते!"}]}'

→ 200 OK   sarvam    in-1   511 ms   ₹0.0021

Normalization

Providers disagree on response schemas, token accounting, and error payloads. OpenLink normalizes all of them into one documented shape before the response crosses back to you. Your code reads one schema, not three.

Regions

Requests are processed in in-1, OpenLink's India region, which is also where Sarvam's low-latency endpoint lives. Additional regions are on the roadmap and will be selected per key.

Data handling

Data retention

By default OpenLink does not persist request bodies. A response may be retried inside the provider TTL window, then discarded. Log retention is configurable from the dashboard.