// the find
higress-group/higress
🤖 AI Gateway | AI Native API Gateway
Higress is an API gateway built on Envoy and Istio, originally developed internally at Alibaba for handling massive-scale traffic. It's now open-source and positioned around AI workloads — proxying LLM APIs, hosting MCP servers, and handling SSE streaming — while still working as a general Kubernetes ingress controller or microservice gateway.
The Wasm plugin architecture is genuinely well-designed: sandboxed, multi-language (Go/Rust/JS), hot-reloadable without dropping connections. This matters a lot for AI traffic where you can't afford reload jitter on long-lived SSE streams. The MCP server hosting through the plugin mechanism is a practical solution — you get auth, rate limiting, and audit logs on your AI tool calls without running a separate proxy. The production pedigree is real: Alibaba runs this at hundreds of thousands of requests per second for actual AI products like Tongyi, not just toy workloads. Config changes propagating in milliseconds (vs. nginx's reload behavior) is a concrete improvement that shows up immediately when you're iterating on routes.
The Chinese-origin bias is unavoidable: default image registry is Alibaba Cloud's Hangzhou region, most of the community activity and docs are in Chinese first, and the enterprise edition push is constant — the README links to Alibaba Cloud's paid product multiple times. If you're not on Alibaba Cloud infrastructure, you'll be fighting defaults. The dependency on Istio is heavy; you're pulling in a lot of control-plane complexity even for simple use cases, which means the learning curve is steeper than Kong or Traefik for teams that just want a straightforward API gateway. The CNCF Sandbox status means it's not yet graduated — governance and long-term maintenance commitments are still being established, which is a real consideration before betting production infrastructure on it.