Introduction

Amux is an LLM API gateway. Requests sent to https://gateway.amux.ai are forwarded to upstream providers according to your configuration, with authentication, billing, routing, and observability unified along the way.

Amux does not train or host models. It focuses on the layer between clients and upstream providers.

The problem#

Integrating one model vendor is manageable. Integrating many multiplies the same work: another SDK, another auth scheme, another invoice, and another availability path to handle.

It is also not a one-off cost. Models change frequently, prices move, and switching models should not require editing code, retesting, and redeploying.

Amux consolidates that layer behind one address.

Design principles#

No markup on inference#

A model is usually offered by several providers at different prices and context limits. We list every one of them in the model catalog and bill at the provider's own price. Discounts we negotiate go into the price you see, not into the middle.

That choice also shapes the rest of the product: if routing does not profit from price spread, there is no reason to steer traffic toward a more expensive provider or to obscure pricing.

Protocols are implemented natively, not translated#

OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and Google Gemini are implemented natively. Tools that support only one of these protocols — such as Claude Code, Codex, or Gemini CLI — typically work after changing base_url.

You can also call across protocols: an OpenAI SDK against Claude, or an Anthropic SDK against Gemini. For what changes in a cross-protocol call, see protocol conversion compatibility.

No silent downgrades#

Unsupported capability combinations fail with a clear error instead of being silently substituted with a near match. This avoids mismatches between what a client asked for and what it actually received.

Your data is yours#

We do not train any model on your requests or responses. Whether request and response bodies are recorded is yours to control, per key or per call. See the privacy policy.

Failures should be legible#

The error body preserves useful upstream detail where possible, including messages such as CUDA out of memory. Attempts that fail during routing retries are not billed.

Each call keeps a full record: inputs, outputs, usage, the provider that served it, and the upstream response on failure.

What we do not do#

Clear boundaries are often more useful than a longer feature list:

• No model benchmarks or leaderboards. Model quality depends heavily on the task, and an inference platform is not a neutral authority for universal rankings
• No prompt hosting, no agent framework. Amux focuses on the infrastructure layer beneath those tools
• No lock-in. Model names use the standard vendor/model format and the protocols are industry-standard, so moving back to direct calls or to another gateway stays straightforward

Where we are#

Shipped: inference endpoints for four protocols, the model catalog with per-provider pricing, pay-as-you-go billing, API keys and a usage dashboard, cross-provider routing with circuit breaking, organizations and workspaces.

Planned work includes subscription billing, broader multimodal support, and cost analysis in the console. This page tracks only capabilities that are currently available.

Where to start#

Start with the quickstart to create a key and make the first request. For team use, see organizations and workspaces; for cross-protocol calls, see protocol conversion compatibility.

The legal texts are the terms of service and the privacy policy.