Roar AI
DocsRequest access
AI Gateway

Every model, on one key

Point your code at one endpoint and reach all of them — chat, embeddings, images, speech and batch. If it speaks OpenAI, it already speaks to us.

THE WHOLE CHANGE
client = OpenAI(
  base_url="https://api.roar-ai.com/v1",
  api_key="roar_live_…",
)
One endpoint, one key, one bill.
55
Models published, each at a rate you can read first
8
OpenAI-compatible endpoints, not just chat
−88%
Median saving on open-source models, blended 3:1 in to out
The difference

The same model should not behave differently on Tuesday

Most gateways pick who serves your request on price, per call. That is why the same model can accept a setting one minute and reject it the next. We choose the route once, and it holds — so what you tested is what ships.

The price holds too. A published rate is stored on the model, so a provider putting their price up costs us margin rather than costing you a surprise mid-month.

Your request
The route we chosePINNED
The same answer, every call
Alternates stay behind it, for when a route fails
The surface

More than chat,
on the same key

Everything below is the OpenAI shape you already write against. Swap the base URL and your existing code keeps working.

Chat completions
/v1/chat/completions
Embeddings
/v1/embeddings
Images
/v1/images/generations
Speech
/v1/audio/speech
Transcription
/v1/audio/transcriptions
Batches
/v1/batches
Files
/v1/files
Models
/v1/models
Control

Give it to the team
without giving up the reins

The part that decides whether a company can actually let people use this: what leaves, what it costs, and who can do what.

Nothing sensitive leaves by accident

Requests are scanned for personal data and secrets before they reach an external model. Redact it, block it, or send that call to a private one instead.

A ceiling you set

Caps on spend per key, per project or per team, counted from real usage as it happens. The tightest one wins and a request over the line is refused.

Keys that only do one job

Every key carries scopes, so the one in a prototype cannot reach the things your production app can.

Your call history, readable

See what was sent and what came back, encrypted at rest and expiring on its own. Turn it off per team if you would rather it was never stored.

Paying less

Three cheaper ways to run the same model

None of these is a different model or a worse one. They are the same call, bought differently.

CACHED INPUT

Repeat a long prefix — a system prompt, a contract, a codebase — and those tokens bill at the cached rate. Nothing to switch on.

BATCH

Send a file of requests, collect results within 24 hours, at roughly half the rate. It runs on the provider’s own batch tier, so the discount is a real saving rather than a gift.

PRIORITY

Faster and steadier on the models that offer it, at a premium. Worth it when somebody is sitting in front of the answer.

Not every model offers every tier. The rate card marks the ones that do.

Agents can use it directly

Point an MCP client here with the same key. It only sees the tools that key is allowed to use.

https://cloud.roar-ai.com/mcp

Change two lines
and try it

We’re open for early access. Bring the code you already have.