One account key is not an access policy
Most teams start AI work by sharing a single provider key. It works until someone needs to know who spent what, a key leaks into a repo, or a customer pastes a card number into a prompt. AI Gateway puts a policy between every caller and every model.
Scoped keys, no sharing
Your account key never leaves the portal. Each team, service, or person gets its own key with its own models, limits and expiry. Revoke one without rotating the rest.
Spend you track
Cap spend per gateway, per person, per key or per end user, over a day, a week or a month. The gateway reserves the cost before a request runs, so a cap is a cap.
Every model behind one URL
Open models on Telnyx GPUs with zero data retention, and your own OpenAI and Anthropic keys, under the same controls. Your code changes its base URL and nothing else.
How control works
Four scopes. Any one of them can say no.
Every request is checked against its key, the person who owns it, the gateway it belongs to and the end user it is for, all at once. If any scope is out of budget, over its rate or blocked, the request stops before it reaches a model.
Gateway
Allowed models, guardrails and a team budget. Every key issued into it inherits the policy. API name: token group
Person
One budget and rate limit across every key a person holds. API name: token user
Key
Can narrow models, add limits and expire. Keys are always capped: a gateway or person can be uncapped, a key cannot. API name: token key
End user
Your customer, named by the user field on each request. Cap or block one customer without touching anyone else. API name: end user
How it works
Change the base URL. Keep your SDK.
Create a gateway, add a policy, issue a scoped key and call any model through the OpenAI SDK. Anthropic models also accept the Anthropic Messages API.
Features
Everything between your code and the model
Set it up in Mission Control under Inference, or drive every part of it from the management API.
Model allowlists
Choose which models each gateway can call, from Telnyx hosted open models to your own OpenAI and Anthropic keys. A key can narrow that list further.
Budgets
Dollar caps per gateway, person, key or end user over a rolling 1, 7 or 30 day window, counted from when you set them. Cost is reserved before dispatch, so nothing overruns.
Rate limits
Requests and tokens per minute on rolling 60 second windows. Callers over the limit get a 429 with a Retry After header.
Guardrails
Ignore, flag or block secrets, card numbers, IBANs, SSNs, emails and phone numbers, separately for prompts and responses, with one policy per gateway.
Bring your own key
Store an OpenAI or Anthropic key once and attach it to any gateway. Billed by your provider, with every gateway control still applied.
Usage analytics
Tokens, cost and status for every request, summarised by gateway, person, key or end user, in Mission Control or through the spend API.
Products
Building AI that reaches beyond the chat?
AI Gateway governs model calls. The rest of the platform runs the code, holds the state and places the call, on the same network.