Side by side.
| Aspect | kibbu | An AI gateway |
|---|---|---|
| Where requests go | Your own hardware first. Cloud is the fallback. | Between cloud model providers, by rules you set: model, cost, latency, availability. |
| API | OpenAI-shaped. Change one URL, keep your client. | Usually one OpenAI-shaped API in front of many providers. |
| Fallback | To OpenAI or Anthropic when the fleet is at capacity or the request needs more than it has. | From one provider to another when a provider fails or is slow. |
| Cost visibility | Every request logged, priced and attributed, local or cloud. | Usually per-request logs, token counts and spend, often with budgets per key or team. |
| Self-hosted models | Serving them is the product: it schedules work onto idle machines and yields to their owners. | Many can point at a self-hosted endpoint. Running that endpoint - machines, models, capacity - is yours to build. |
Pick a gateway when.
- All of your traffic belongs on cloud models, and the job is choosing between providers.
- You need failover between providers, caching or per-team budgets across many vendors.
- A human is waiting on most of your requests, and latency matters more than cost.
Pick kibbu when.
- A large share of your AI traffic is agents, evals, embeddings or nightly jobs that nobody is waiting on.
- The company already owns machines that sit idle for part of the day.
- You want that work on hardware you already own, with cloud kept for what needs it.
Moving one workflow takes three changed fields. The quickstart shows them, and kibbu.md lists when kibbu is the wrong choice.
Questions.
Is kibbu an AI gateway?
No. A gateway's job is routing between other people's clouds. kibbu serves requests on your own hardware first, and routes to OpenAI or Anthropic only for what that hardware cannot serve.
Can I keep my gateway and use kibbu?
kibbu is an OpenAI-compatible endpoint, so a gateway that accepts one can treat kibbu as another provider. Check your gateway's docs for how it adds a custom OpenAI-compatible base URL.
Does a gateway lower my token bill?
A gateway can send a request to a cheaper model or answer it from a cache. Whatever reaches a provider is still billed per token. kibbu serves suitable work on hardware you already own, so it never reaches a provider's per-token bill.
What does kibbu do with requests my machines cannot serve?
It sends them to OpenAI or Anthropic, priced, logged and attributed, and your application sees the same API either way.
