ML Junction

About ML Junction

ML Junction is an API gateway that routes requests to the right AI model across multiple providers. It offers transparent pricing at provider rates plus one routing fee, whether using prepaid credits or your own provider keys (BYOK). The platform features adaptive models that can switch between candidates based on performance, automatic failover when providers are busy or down, and detailed request receipts showing exact routes, costs and privacy policies. It supports the OpenAI and Anthropic SDKs by changing only the base URL and API key, and provides governance features including spend caps, rate limits, and IP-based access controls.

• transparent pricing
• model abstraction
• ai routing
• multi-provider
• openai compatible
• api gateway

FAQ

Do I have to change my code to use ML Junction?

Only the base URL and the key. The OpenAI and Anthropic SDKs work as they are, and the native /v1/responses API adds routing, receipts and privacy controls when you want them.

Does routing through ML Junction break prompt caching?

No. A conversation stays on the route and account that served its previous turn while that route is healthy, so the provider cache is still there on the next turn. Every response reports whether the cache hit and why.

What happens when a provider fails?

The request moves to the next permitted route at once; only a brief blip is retried in place. A provider failing most of its calls is benched for everyone until it recovers. Every attempt is named in the request trace.

How does an adaptive model decide to switch?

A challenger has to clear the gates you set: enough samples, reliability floors, zero critical failures, a minimum improvement and a cooldown for scheduled evaluations, and stage-by-stage evidence with latency and error guardrails for production rollouts. Each change creates a revision you can roll back.

Does ML Junction train a router on my data?

No. Candidates compete on your own cases and outcomes through controlled trials and ordinary statistics. There is no classifier learning to predict a winner from your prompts.