Why I'm building Busbar

Before I write a line of code, I want the “why” on the record. There’s no repo yet, nothing to install, just a conviction about the piece of infrastructure the multi-model era is missing, and a description of what I’m setting out to build.

The problem

Every serious AI application is becoming multi-model and multi-provider. You want Claude for one thing, GPT for another, a cheap model for classification, and a fallback when your primary is rate-limited. The moment you have more than one provider, something has to sit between your app and all of them. That something is a control plane.

The gateways that exist today mostly do two things. They flatten every provider to OpenAI’s shape, and they retry on failure. Both are quietly lossy.

• Flattening to OpenAI throws away what makes each provider worth using: Anthropic’s thinking blocks, Gemini’s safety settings, Bedrock’s tool-use envelope. You get portability by giving up capability.
• Retry-on-failure is error handling, not failover. The call throws, and then something retries, after your user already felt the stall. And a naive retry can hammer a provider that’s already down.

Neither is wrong, exactly. They’re just not enough once the control plane is load-bearing infrastructure.

What I’m building

Busbar is the reliability and fidelity layer for multi-model AI traffic:

• Lossless translation, both ways. Speak any of six wire protocols in, any provider out, through a superset intermediate representation, so native features survive the hop instead of being flattened away.
• Failover inside the request. Reroute across providers before the client sees a byte, even mid-stream, within a deadline and hop budget. The user never sees the stumble.
• A circuit breaker that knows whose fault it is. Classify each failure (provider outage, your bad request, context-length, auth or billing) and treat each differently, instead of retrying into a wall.
• One static Rust binary. No Python sidecar, no interpreter, no GC in the request path. Your keys, your network, your data path.

Why me, why now

Multi-model is going mainstream this year, and the control plane is where the reliability and portability promises get made or broken. I think the right architecture is reliability-first and lossless, not OpenAI-shaped and retry-based, and I’d rather build that from the metal up than bolt it onto an interpreted proxy.

It’s day zero. Nothing exists yet, not even a repo. But I’m building this in public, and I’ll post every milestone here as it lands. If any of this resonates, if you’re running multi-provider LLM traffic and feeling these edges, I’d love to hear from you.