HotRiceNoodles / CrossLink

What is CrossLink?

CrossLink is a Go-based API gateway that sits between your application and multiple LLM providers. It routes requests to OpenAI, Anthropic, Azure, DeepSeek, and other compatible services while handling load balancing, automatic failover, rate limiting, and cost tracking in one place. You point your existing OpenAI SDK at it, and CrossLink handles which provider actually handles the request, retries on failure, and logs usage. It also includes an MCP gateway for Model Context Protocol tool discovery and a Vue admin dashboard for configuration.

Why is it gaining traction?

The dual-API support is the hook: you get both OpenAI and Anthropic endpoints from a single gateway, which matters if you're mixing models or migrating between providers. The built-in circuit breaker and retry budget prevent cascading failures when a provider goes down. Redis caching with gzip compression reduces costs on repeated queries. The Chinese national cryptography mode (SM2/SM3/SM4) is unusual in open-source gateways and addresses compliance requirements for deployments in regulated markets. Multi-instance sync via Redis Pub/Sub means you can scale horizontally without losing routing consistency.

Who should use this?

Backend teams managing multi-provider LLM infrastructure who need failover without building it themselves. DevOps teams requiring per-key rate limits, budget caps, and usage analytics across teams. Organizations with Chinese compliance requirements that need GM cryptography support. Teams running MCP-based agents who want a managed proxy layer with permission controls.

Verdict

CrossLink is a serious project with production-ready deployment configs, test coverage, and a well-documented feature set. At 18 stars, the community is tiny and the credibility score of 1.0% reflects that early-stage risk. The code quality and breadth of features suggest the maintainers know what they're doing, but a project this young with this few watchers could change direction or abandon ship. Evaluate it for non-critical internal tooling where you can absorb that risk, but hold off on production customer-facing deployments until the community grows.