Production AI integrations, explained

One endpoint in front of several model providers — routing rules you can change without a deploy, a fallback path that is part of the design, and cost measured per request. Built for the products we run ourselves.

Get API access How it works

Private beta — no self-serve sign-up, no published SLA. We say so here rather than after you have filled in a form.

1,405,091

requests in a week

through our own gateway, 09.09 to 15.09.2026

2.8 s

median answer

half are faster than this

26.3 s

nine in ten answers faster than

speech and image profiles included

0.16%

requests got no answer at all

another 10.31% was saved by the fallback path

Measured 16.09.2026 from our own gateway log in a week. One gateway, twelve task profiles together — from short checks to speech recognition and synthesis. The profiles differ, so the median and the tail differ by a lot.

Gateway is responding

checked at 20:16 UTC

The single endpoint in front of our products is accepting requests. The time of the check is next to the heading: this page does not poll the gateway in your browser and does not pretend that it does.

Try it here

POST /api/v1/probe

Your text goes through the same gateway our products use. Which model answers is decided by a routing rule, not by this page.

Do not enter personal, confidential or client data.
Your text is sent to an external model provider. We do not keep your prompt, but the model's answer stays in our request log for 7 days — we need it to investigate failures. How this works — in the privacy policy.

Or try an example:

Up to 600 characters, no sign-up.

Other ways to solve this

Approach	Who it suits	What it costs you
Call each provider directly	One provider, one use case, and no plan to change either	The provider outage becomes your outage, and switching later means touching every call site you wrote
A routing library in your codebase	You want provider abstraction and are happy to own it	Routing rules ship with your app: changing a model means a deploy, and cost stays invisible until the invoice arrives
A commercial gateway product	Standard needs, budget for per-seat or per-token pricing, data allowed to transit a third party	You configure what the vendor anticipated; anything else waits for their roadmap
QueryPort	You need routing you can change without a deploy, a real fallback path, and cost measured per request	We are small and the gateway is in private beta: no self-serve sign-up, no public SLA yet, and access is granted case by case

When we are the wrong choice

Being straight about the edges: in these cases another route is the better one.

You want a self-serve API key right now

There is no public sign-up, and we are not pretending otherwise. Documentation, usage terms, rate limits and support are not at a level we would call a product yet.

What to do instead: Talk to us if the fit is close, or use a provider directly until we open up.
You need a published SLA and enterprise procurement

We run this for our own products first. Uptime is real but not contractual, and we will not sign a number we cannot yet defend with a year of data.

What to do instead: Pick a vendor whose SLA is part of the product. That is a legitimate requirement, not a formality.
Everything must stay inside your own network

The gateway talks to external providers. A fully local setup is different work: your own hardware, your own weights, your own operations.

What to do instead: Worth a separate conversation. Local models solve the data question and lose on quality and cost.
You are looking for a model, not an integration

We do not train models. We route to them, and we care about what happens when routing goes wrong.

What to do instead: Go to the providers directly, then come back when the operational questions start.

Glossary

Gateway: A single endpoint between your code and several model providers. Your application talks to one address; which model answers is decided by a routing rule you can change without redeploying the application.
Routing profile: A named rule set per task type — text, OCR, vision, speech. Each profile has its own provider order, its own limits and its own cost per request.
Fallback path: What happens when the primary provider errors or goes quiet. Without one, a provider incident becomes your product incident, usually at peak hours.
Token: The unit providers bill in, counted across both your prompt and the answer. This is why the length of a system prompt is a line item, not a formatting detail.
Time to first token: Delay between the request and the start of the answer. For an interface it matters more than total time: an answer that has started reads as fast even when it finishes slower.
Prompt injection: Text arriving from a user or a document that carries instructions for the model and displaces yours. Mitigated by separating roles and validating what came back — not by trusting the model.

Questions we get

Can I get an API key today?

Not through a sign-up form — there is none. Access is granted case by case while the gateway is in private beta. Tell us what you are building and we will say plainly whether it fits today or not.

What does the gateway actually do?

Your application sends a task profile — text, OCR, vision, speech — rather than a model name. The gateway picks the provider by rule, retries elsewhere on failure, and records latency, cost and outcome per request. Changing the rule does not require a deploy on your side.

Why send a profile instead of a model name?

Because models get retired. A model name hardcoded in an application turns into a failure with an unhelpful error message the day the provider removes it. A profile survives that: the rule changes, the application does not.

Which providers do you support?

Several, and the list changes as providers change their terms and pricing. We name them during onboarding rather than on this page: a list published once and left alone is worse than no list.

What happens to the data passing through?

Request and response text transits the gateway. What is logged and for how long is agreed before any work starts, and provider-side training on your data is a setting we check explicitly rather than assume.

Is there vendor lock-in?

One endpoint and one key. Removing the gateway means pointing your client at a provider directly — no rewrite of your application. We would rather you could leave than have you stay because leaving is expensive.

What is your uptime?

The block above checks the gateway live every time this page is opened, and it reports failures as failures. We do not publish an uptime figure yet because we do not have a year of data to stand behind.

Do you also build integrations?

Yes — that is the larger part of our work, and it is where the gateway came from. The engineering side is described on our Russian-language site; if you need the work rather than the endpoint, write to us and we will answer in English.