Overview
One self-hosted endpoint in front of every model a developer uses, so existing OpenAI-compatible clients keep working while routing, fallback, and spend stay under their control.
The hard part
Failing over between upstream tiers without corrupting a streamed response, and reporting token savings as measured numbers rather than a saver’s own estimate.
The approach
The executor peeks the first chunk before committing to a tier and walks ordered combos transparently; savers run a deterministic pipeline that fails open and is re-measured from rewritten messages, and the Vue dashboard is compiled into the binary with its SQLite migrations.
