Blog
Notes from the team on AI control-plane reliability, fidelity, and performance.
Context compression belongs on the request path, not bolted onto your app. Busbar runs it as a rewrite gate: prompt text in, a smaller body out, before routing and before dispatch. And because the hook self-describes its settings and self-reports its metrics, you configure it and read its savings entirely through the frozen Admin API — no second dashboard.
Hooks put your logic on the normalized request path across all six protocols. Auth becomes a pluggable chain you can compile out. And the admin API v1 is frozen: the surface tools get to build on.
Point Claude Code at Busbar with one environment variable — get observability, failover, and budgets for the agent you already use. Then the twist: because Busbar translates between protocols, you can point Claude Code at a Gemini or Bedrock model instead.
Everyone asks for automatic model selection by task, latency, quality, and cost. It is not a product, it is a hook. Busbar runs yours two ways: a compiled Rust binary on a local socket that decides in about 8 microseconds, or a webhook in any language. Both wired to the same failover and fail-safe machinery.
Embeddings, moderations, image generation, and audio (transcription and speech) now run through Busbar, and every one is cross-protocol. A Gemini client can call embeddings on Bedrock, an OpenAI client can route images and audio to Gemini, and every answer comes back in the caller's own dialect. Lossless, both ways.
Why a control plane earns its keep with one provider: key custody, hard caps, and real metrics on day one, plus an option you can exercise mid-incident.
A fast, lightweight, single-binary AI control plane isn't my roadmap. It's what shipped. Straight answers on memory, latency, throughput, reproducibility, and what you actually deploy.
The HTTP API, config schema, and six wire-protocol contracts are now frozen under Semantic Versioning, hardened across a multi-round security and correctness audit. It's production-ready.
From an empty repo to a feature-complete, API-stable release candidate. Six protocols, lossless translation, in-flight failover, and fault-attributed breaking, in one Rust binary.
As teams go multi-model, the control plane becomes critical infrastructure, and 'flatten everything to OpenAI and retry on failure' isn't enough. Before I write a line of code, here's what I'm setting out to build.