+
+
+
+
Cencori for Developers ·150+ models·10 providers·1 SDK

The AI cloud infrastructureto run, secure, andscale AI products.

For developers, AI engineers, and research teams shipping real products. One SDK to every frontier model. A compute layer to train, host, and deploy your own. Regional cloud you can pin workloads to.

+
+
+
+

Ship in five lines

Run.

OpenAI-compatible request shape. If you have OpenAI code, changing one import gets you multi-provider routing, caching, logging, and spend caps — without touching business logic.

TypeScript
import { Cencori } from '@cencori/sdk';

const cencori = new Cencori({ apiKey: process.env.CENCORI_API_KEY });

const response = await cencori.chat.completions.create({
    model: 'gpt-4o',
    messages: [{ role: 'user', content: 'Explain hedging in one sentence.' }],
});

console.log(response.choices[0].message.content);
+
+
+
+

The platform

One platform. Three layers.

Cencori is not a gateway. It is a cloud — delivered in three composable layers so you can adopt what you need today and grow into the rest.

+
+
Live

Cencori Gateway

One SDK to every frontier model. Chat, vision, documents, embeddings, agents — with governance, caching, and observability on by default.

  • 150+ models across 10 providers
  • OpenAI-compatible API
  • BYOK & multi-provider fallback
Q4 2026 · Early access

Cencori Compute

For research teams and AI engineers who need to train, host, and deploy their own models. Point at a repo, wire a dataset, benchmark, publish — your model becomes an endpoint in your gateway.

  • Repo → dataset → training
  • Benchmarks against reference models
  • One-click publish to the gateway
2027 · Preview

Cencori Cloud

Region-native inference and workload hosting. Pin a project to a region and enforce residency at the platform layer — Lagos colocation partnerships already in flight.

  • Region-locked storage & inference
  • Residency by contract
  • CPU + GPU hybrid capacity
+
+
+
+
TypeScript
// Same code. Any provider. Zero rewrites.

await cencori.chat.completions.create({ model: 'gpt-4o', messages });
await cencori.chat.completions.create({ model: 'claude-sonnet-4-6', messages });
await cencori.chat.completions.create({ model: 'gemini-2.0-flash', messages });
await cencori.chat.completions.create({ model: 'mistral-large-latest', messages });
await cencori.chat.completions.create({ model: 'grok-4', messages });

Multi-provider

Any model. One line.

Switch providers by changing a string. Set automatic fallbacks so a provider outage never becomes your outage. When your own fine-tune ships from Cencori Compute, it becomes just another model in the same list.

+
+
+
+

@cencori/react

We ship the UI too.

Drop-in Chat, VisionUploader, DocumentReader, and VoiceCall components. Streaming, accessible, themable. You can put an AI surface in front of a user tonight.

React
import { Chat, VisionUploader } from '@cencori/react';

export default function App() {
    return (
        <>
            <Chat model="gpt-4o" apiKey={process.env.NEXT_PUBLIC_CENCORI_KEY!} />
            <VisionUploader onExtract={(text) => console.log(text)} />
        </>
    );
}
+
+
+
+

Capabilities

Everything you need to ship AI.

One SDK covers the entire surface. Adding a capability is a method call — not another vendor account.

Live

Chat & tool use

Streaming, structured outputs, JSON mode. Every provider through one shape.

Live

Vision

Image analysis, OCR, multi-image reasoning in a single call.

Live

Documents

Extract, summarize, and query PDFs. Native parse first, LLM second.

Live

Embeddings

Portable vectors across every major embedding model.

Live

Agents

Multi-step tool use with retries, timeouts, and cost caps built in.

Soon

Voice

TTS, STT, and realtime bidirectional voice. Multi-provider.

Q4 2026

Fine-tuning

Train and host your own models. Ships with Cencori Compute.

Live

React components

Drop-in Chat, VoiceCall, VisionUploader, DocumentReader.

+
+
+
+

Get started

Ship AI today. Train your own tomorrow.

Grab an API key, paste it into your existing code, and you’ll have a first response in about ninety seconds. When you’re ready to train your own model, Compute is a queue away.