promptguard-sdk/integrations/vercel-ai, which scans each prompt before the model is called and, if you ask it to, each response. Its setup, options and streaming behaviour are documented in the Node.js SDK guide.
Use it rather than init(). The Vercel AI SDK’s providers, such as @ai-sdk/openai, make their own HTTP calls instead of using the openai package, so auto-instrumentation does not see them.
The proxy works too: providers created with a custom baseURL and headers send their requests through PromptGuard like any other client.