One URL. A cap that's already on.
Change base_url in the OpenAI SDK you already use and you're on the gateway: Claude with nothing to set up, and any other provider through your own key or your own endpoint — behind one key, with a spend cap per project enforced before the call goes upstream, not a limit you discover on the invoice.
$15 a month, and it isn't just this app — every Parsons product is on the same subscription. Cancel any time. The cap is on before you ask — $5 a month per project by default, checked against a strongly-consistent counter before every upstream call.
Change one line. Keep the SDK.
Point the SDK here.
base_url and a key. That's the migration. The OpenAI SDK you already use, unchanged.
Give me a key for the invoice app.client = OpenAI(
base_url="https://ai-gateway.parsons.ai/v1",
api_key="sk_live_…",
)Paste one link into your AI.
Your account comes with one private address. Paste it into Claude, ChatGPT or Cursor once, and AI Gateway's tools are there. No AI of your own? The console has chat built in.
https://mcp-hub.parsons.ai/p/prof_••••••••/mcpPick a model per call.
Six Claude models are managed — priced in the open, nothing to set up. Everything else on earth reaches the same URL through your own provider key or your own endpoint, logged the same way.
Route the classifier to the cheapest model that clears 90 on the eval.Watch the spend, and the cap.
Usage by key, by model, by day. A cap per project, enforced upstream of the call. Raise it when you mean to.
How much has the invoice app spent this month, and on what?A gateway that prints its prices and stops at your number.
Managed Claude, and everything else through your key
Six Claude models with nothing to set up. Any other provider through your own key, or your own endpoint, on the same URL — same logs, same cap.
A cap that's enforced, not reported
Per project, in dollars, on by default at $5. Checked against a strongly-consistent counter before the upstream call. Not an alert — a wall.
Keys per project, per environment
Each key's spend is its own line. Rotate with an overlap. Revoke and the traffic stops.
Your provider keys, or ours
Bring a key and calls route through it at the provider's price. Use ours and pay tokens at cost and a half, prepaid.
Logs you can read
Every call: model, tokens, latency, cost, which key. Ask your AI why yesterday was expensive and it reads the same rows.
Models & pricing, asked
The console's own chat knows the price list and your logs. It answers the question you'd otherwise open three tabs for.
It routes your calls. It cannot invoice you afterwards.
- ×Spend past the capThe cap is checked before the upstream call. Past it, the call is refused with a clear error, not billed.
- ×Invoice youManaged spend draws from prepaid credits. Nothing arrives after the fact.
- ×Read your prompts to train anythingRequests pass through to the provider and are logged as metadata — model, tokens, latency, cost. Content is not retained.
- ✓Route any OpenAI-compatible call — managed Claude, or any provider through your key
- ✓Issue keys, set caps, bring provider keys
- ✓Read usage and logs, by key, model and day
There is no security slogan on this page and no badge we haven't earned. Everything in that list is a thing you can check; the full mechanism and every tool are on the details page.
Fifteen dollars a month — and it isn't just this app.
Is it worth it? Start where we look worst.
Going direct to one provider is cheapest if one provider is all you'll ever use. Here's what the gateways publish.
Ours is $15 for the gateway with the cap built in, and the same $15 is Auth, Database, Deploy and the rest the app needs. Managed tokens cost 1.5× — use your own keys and they cost what the provider charges. And there is no free tier because a free gateway is one that learns from your prompts.