ngrok's AI Gateway Routes to Private Models Without Exposing Them

One key and one URL route across hosted providers and self-hosted models, with access control, observability and automatic failover between models or keys. Local Ollama and vLLM endpoints sit alongside cloud providers, reached through ngrok’s existing tunnel network rather than the public internet. Model gateways are thoroughly commoditised — Vercel, OpenRouter, LiteLLM and Cloudflare all ship one — so the routing layer is not the differentiator. The tunnel network already terminating inside customer VPCs is. Copying the gateway takes a weekend; copying the network underneath it took a decade. New accounts get $1 in credit at app.ngrok.ai.