Cencori for Developers

Cencori is where you build AI products.

Not a model you call and then go stitch together ten other services around. One platform, one SDK — and everything a real AI product needs, waiting for you as a method call instead of another vendor account.

Here is what that actually means. Not a feature list — a list of things you can go build today.

import { Cencori } from 'cencori';

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

const res = await cencori.chat.completions.create({
    model: 'gpt-4o',
    messages: [{ role: 'user', content: 'Ship something today.' }],
});

That is the whole setup. Everything below is one import away from that line.

• Chat with any provider, switch with a string. OpenAI, Anthropic, Google, xAI, Mistral, DeepSeek, Meta, Groq, Cerebras, Perplexity, Cohere — one API shape. Change 'gpt-4o' to 'claude-sonnet-4-6' and you are on a different model. Nothing else changes.
• You already have the code. If your app talks to OpenAI or Anthropic today, you are one import away — keep your model, gain routing, logging, security, and spend controls without touching business logic. Most migrations are a single line.
• Stream tokens for responsive UIs, call your own functions with tool use, and get type-safe, validated JSON back with structured output — no more parsing model prose.
• Moderate content through a dedicated safety endpoint before it reaches a user.
• Use Cencori from any language. Point an OpenAI-compatible client in C#, PHP, Ruby, Java — anything — at Cencori by changing one base URL. No SDK required.

• See images. Analyze, describe, OCR, and classify — ship a receipt-to-JSON extractor, an ID reader, or a "take a photo and ask about it" feature in a single call.
• Read documents. Extract, summarize, and answer questions over PDFs. Native text extraction is free — you only pay tokens on the actual questions, so a contract Q&A tool costs almost nothing to run.
• Give any product a voice. Text-to-speech and speech-to-text — take a voice note, including Yoruba, Hausa, and Igbo, and get back a clean transcript with speaker labels.
• Generate images across models, and create embeddings for search and retrieval — all through the same key, the same SDK.

• Let your app remember. Give any conversation persistent, per-user memory — so your assistant knows who it is talking to before it replies, and remembers what mattered after.
• It is not a raw vector store you babysit. Cencori extracts what is worth keeping, ranks what to recall, resolves entities into a graph you can query, tracks what is still true over time, lets stale facts decay, and redacts PII on the way in — automatically, isolated per organization, scoped by namespace.
• Upload a file straight into memory, or run full RAG that pulls the right context into a prompt and hands you back the sources it used.

• Build agents that survive production. Durable sessions that keep running across turns, survive crashes, and resume exactly where they stopped — with no idle compute cost while they wait.
• Put a human in the loop. An agent can pause for approval before it does something consequential, then continue the moment someone signs off.
• Give agents real tools. One Responses API endpoint gives you web search, file search, a code interpreter, and function calling — the building blocks of an agent that actually does things.
• Scope keys to an agent, so what an agent can reach is bounded by design.
• Bring your own framework. Power CrewAI, AutoGen, LangChain, and LangGraph with Cencori underneath — same routing, security, and observability.

• Redact PII in the request path, before it ever reaches a provider — emails, phone numbers, card data, account numbers, even obfuscated formats.
• Block jailbreaks and prompt injection as they arrive, and filter harmful output on the way back out with thresholds you set.
• Write your own data rules — keyword, regex, JSON-path, or model-based — that block, redact, mask, or tokenize whatever you decide can't leave your boundary.
• Enforce policy and keep an audit trail. For teams that need it, every AI decision can run against policy in the request path and land in an immutable, verifiable ledger — the difference between "we think it's safe" and "here's the proof."

• Stop paying for the same answer twice. Exact and semantic caching reuse responses for repeated and similar prompts, cutting cost and latency automatically.
• Never let a provider outage become your outage. Automatic failover and circuit breaking route around a down provider before you notice.
• Bring your own keys, or plug in your own self-hosted models as a custom provider — routed, secured, and observed exactly like the built-in ones.

• See everything. Every request logged — cost, tokens, latency percentiles, provider and model breakdowns, geography, failovers — with search and export. Pull it into your own dashboards through the Metrics API, and browse every available model through the Models API.
• Manage keys and environments — secret, publishable, and test keys, with domain restrictions — so client-side and server-side, staging and production, stay cleanly separated.
• Version your prompts in a registry, deploy them, and track how each one performs over time.
• React to events. Subscribe to signed webhooks and fire your own workflows when things happen.
• Set rate limits and spend caps so a runaway loop never becomes a runaway bill.

• Put an AI surface in front of a user tonight. First-party React components — chat, voice, and vision uploaders — that stream, handle errors, compress and validate uploads, and theme to your app.
• Keep the tools you love. First-party providers for the Vercel AI SDK and TanStack AI work with streamText, useChat, and the rest, out of the box.

• Bill for the AI you build. End-user billing lets you put rate plans and quotas on your customers, meter their usage, generate invoices, and collect through Stripe Connect — the metering and payouts are handled, so you can run a business on top of Cencori, not just a project.

• SDKs in TypeScript, Python, and Go — typed responses, streaming, typed errors (AuthenticationError, RateLimitError, SafetyError), built-in retries, and edge-runtime support.
• Start in one command with create-cencori-app, and expose the whole thing to agents through the MCP server.
• Go no-code with n8n, Zapier, and Make, or wire straight into Supabase, Neon, and Firebase.
• Check what you shipped. Cencori Scan reviews your repo for leaked secrets, PII, exposed routes, and vulnerabilities — and writes AI fixes right on your pull requests.

We ship in the open, and the platform grows underneath your feet without breaking your code. Coming next: Compute — train, host, and deploy your own models until your fine-tune is just another entry in the same model list — and Cloud, region-native inference you can pin a workload to.

You do not have to adopt all of it. Change one import and you have multi-provider routing, logging, and spend caps. Add memory when you need state. Add agents when you need action. Add billing when you have customers to charge. The surface widens; your code keeps working.

Grab a key, paste it into the code you already have, and you'll have a first response in about ninety seconds.

Build different.

Start building → · Read the docs →