An AI gateway from ngrok, Inc., the company known for its developer-facing tunneling and ingress platform. Cloud models from OpenAI, Anthropic, and Google, along with local models you run on your own machine or server, are all reachable through a single endpoint at gateway.ngrok.ai with one access key. Your application no longer needs to care which provider it is talking to: failover during provider outages, key management, and usage visibility are all handled at the gateway. After an initial early-access release, it was relaunched as the new ngrok.ai with a dedicated dashboard and API.
Key Features
- One URL, any SDK: Point your request baseURL at
https://gateway.ngrok.aiand pass an ngrok access key. Existing code written against the OpenAI SDK, Anthropic SDK, or Vercel AI SDK works largely unchanged - Local models through the same endpoint: Connect to models you host yourself with Ollama or vLLM without provisioning a public IP or opening ports. ngrok’s own tunneling technology is the foundation
- Automatic failover and retries: If a provider or key stops responding, the gateway moves on to the next model or the next key in the order you configured. There is no need to add error-handling code on the client side
- Scoped access keys: Issue keys at a granularity like “this app may use only this model from this provider.” The real provider API keys stay consolidated at the gateway instead of being distributed to apps and developers
- Usage and cost visibility: A dashboard shows token consumption, latency, and errors, and cost can be broken down by provider and by model
- Configurable via API, CLI, and Terraform: Everything is configurable through the API, so it can be managed declaratively from a CLI, coding agents, or Terraform
Pricing
Use of the AI gateway itself is metered against prepaid credits. Every successful request through the gateway carries a processing fee of $0.05 per million tokens, on top of inference cost. Bringing your own provider keys (BYOK) is free on all plans, and the provider bills you directly for model usage. If you use the OpenAI and Anthropic inference built into ngrok, model cost is deducted from your credits as well. Credits start at a $5 minimum purchase and expire one year after purchase.
Account plans (shared with the tunneling platform) come in four tiers.
| Plan | Monthly price | What’s included |
|---|---|---|
| Free | $0 | $5 one-time credit, up to 3 endpoints, 20,000 requests per month. AI gateway credits can also be purchased |
| Hobbyist | $10 (annual billing saves $24 per year) | $10 in credit each month, up to 3 endpoints, 100,000 requests per month |
| Pay-as-you-go | $20 plus usage | $20 in credit each month, unlimited endpoints |
| Enterprise | Contact sales | SSO / SAML, dedicated support, HIPAA and SOC 2 |
Pricing is current as of August 2026. Check the official site for the latest pricing.
Pros & Cons
✅ Pros
- Collapses per-provider SDKs and keys into a single base URL and key
- Provides failover against provider outages and rate limits without rewriting application code
- Places cloud models and local models behind the same interface, making it easy to switch based on cost or confidentiality
- Scoped keys mean production API keys never have to be handed around the team
- You can start without a monthly subscription simply by purchasing credits
⚠️ Cons
- Routing through a gateway adds a processing fee and one more network hop
- It adds a point of failure, so an outage of the gateway itself has broad impact
- Credits are prepaid and expire after a year, so balance management is needed until your usage is predictable
- It is a relatively new service, and supported providers and features may keep changing
- Some capabilities, such as team features and internal endpoints, assume a paid plan
Comparison with Similar Services
| Criteria | ngrok.ai | OpenRouter | LiteLLM | Cloudflare AI Gateway |
|---|---|---|---|---|
| Delivery model | Managed gateway | Managed router | Open-source proxy (usually self-run) | Managed gateway |
| Local model access | Supported via tunneling | Generally not supported | Supported (self-hosted) | Limited |
| Failover | Supported | Supported | Supported | Supported |
| Scoped key issuance | Supported | Supported | Supported | Depends on Cloudflare’s auth mechanisms |
| Billing model | Prepaid credits plus token fee | Fee on credit purchases | Software is free (you bear operating cost) | Free tier available |
Being able to attach local models to the same gateway without exposing them is ngrok.ai’s biggest differentiator. Conversely, if the breadth of the model catalog and routing options matters most, OpenRouter is the candidate; if you want everything closed inside your own environment, self-running LiteLLM is.
Who Is It For
- Developers juggling several LLM providers whose key management and switching code has grown messy
- Product teams that want to keep a production service running through provider outages and rate limits
- Teams that want confidential workloads on local models and everything else on cloud models
- Administrators who need to restrict which models each app or developer may use, and track AI cost by team
- People already using ngrok for tunneling who want their AI traffic on the same operational footing
Summary
ngrok.ai is a gateway for pushing the operational concerns of LLM usage ─ where to connect, how to distribute keys, what to do when something fails ─ out of the application and into one place. The ability to include local models behind the same endpoint without exposing them is a strength that comes naturally to a company built on tunneling infrastructure. A reasonable way to start is to swap only the baseURL of an existing app, using the free tier and a small credit balance, and see whether the failover and visibility hold up in real operation.