// the find
looplj/axonhub
⚡️ Open-source AI Gateway — Use any SDK to call 100+ LLMs. Built-in failover, load balancing, cost control & end-to-end tracing.
AxonHub is an AI gateway written in Go that lets you point any OpenAI, Anthropic, or Gemini SDK at a single endpoint and route calls to whatever provider you actually want, with failover, load balancing, RBAC, and per-request cost tracking built in. It targets teams running multiple LLM providers who are tired of maintaining separate integrations or who want budget guardrails and request-level observability without rolling their own. If you've ever had to refactor a codebase to swap from GPT-4 to Claude mid-project, this is what you needed.
The zero-code-change migration story is real and well thought out — changing one base URL is genuinely all you need to switch providers for most SDK users. The ent ORM plus GraphQL stack is a solid, typed foundation that avoids the raw-SQL sprawl most gateway projects end up with. Multi-database support (SQLite for local, Postgres or TiDB for production) with automatic migrations means the deployment story scales from a single developer to a proper production setup without rewriting config. The tracing and per-request cost breakdown fills a real gap — most teams running LLMs have no idea what individual requests actually cost until the bill arrives.
The dual Apache-2.0 / LGPL-3.0 license split is unexplained in the README, and that ambiguity is a blocker for anyone in a legal-review-heavy org — you have to dig into the LICENSE file to understand what's covered by which. AWS Bedrock and GCP Vertex are still in testing status, which matters because those are exactly the providers enterprise teams with compliance requirements actually use. The SQLite default is fine for local dev but the README doesn't warn clearly enough that it's not suitable for anything concurrent — someone will deploy it with SQLite in production and wonder why things break under load. Realtime API support is a TODO, which is increasingly table stakes as more workloads move to streaming voice and live agents.