Confidential AI API: Secure LLM Inference at Scale

In one sentence: Prem API provides OpenAI- and Anthropic-compatible model routes with client-side encryption and confidential-computing evidence. Review the current measurement-policy and route-assurance gaps before making a workload-specific guarantee.

The Problem

When you use a standard AI API, your prompts, files, and conversations are processed on infrastructure that you do not control. HTTPS protects data in transit, but the service normally receives plaintext for processing. For many teams, this model is acceptable. Teams handling patient records, financial data, legal documents, source code, or trade secrets may need a narrower processing boundary and evidence they can inspect. Prem API changes where plaintext is available and adds confidential-computing evidence. It does not remove every dependency or trust assumption.

What Prem API Does

Prem API exposes supported chat and audio routes. The TypeScript SDK or local Confidential Proxy encrypts the request payload before network egress. The gateway handles ciphertext and operational metadata, and the selected processing runtime decrypts the payload for inference.

Client-side encryption

The SDK or local proxy encrypts the request payload before sending it to the Prem gateway.

Gateway boundary

The gateway receives payload ciphertext plus authentication, model, size, timing, routing, and billing metadata.

Protected processing

Supported confidential routes process plaintext inside a Trusted Execution Environment (TEE). Attestation evidence and policy define what the client can conclude about that runtime.

The Comparison with a Standard AI API

	Standard AI API	Prem API
Access to your data	The provider, the provider’s staff, and possibly the cloud host	Your application and the protected processing runtime; a local proxy also sees plaintext when used
Gateway payload	The service normally receives the request body	The Prem gateway receives ciphertext plus operational metadata
Processing boundary	Provider-managed workload	A supported confidential route uses a protected runtime; verify the selected route
Security evidence	Provider policy, controls, and audit material	Hardware-signed evidence plus measurement, TCB, and route policy
Host access control	Defined by the provider’s architecture	Confidential-computing hardware is designed to restrict host access; deployment validation still applies
API compatibility	Provider-specific	Supported subsets of OpenAI- and Anthropic-compatible interfaces

Who Prem API Is For

Regulated Industries

Healthcare, finance, legal, and government teams that need to map an AI data path to their own legal, compliance, and security controls.

AI Application Builders

Development teams that need a client-encrypted inference path and can validate the supported API surface.

Security-First Organizations

Teams that require hardware evidence and an explicit policy decision in addition to provider documentation.

Enterprises with Sensitive IP

Companies evaluating AI for internal documents, proprietary code, or strategic planning under a defined threat model.

What You Can Do Today

Prem API exposes a supported subset of OpenAI-compatible and Anthropic-compatible interfaces:

Chat with AI models: Streaming conversations, multi-step reasoning, and tool use
Transcribe audio: Use the currently enabled Deepgram model and its response shape

The Prem API has two inference modes:

Confidential inference: Client-side encryption, a Trusted Execution Environment, and attestation evidence that you can examine. Use this mode for sensitive data and for production workloads.
Zero Data Retention: An OpenAI-compatible endpoint that needs no code changes. Prem sends the request to a trusted partner, and the partner keeps no content. Use this mode for evaluation, for prototypes, and for public data.

See ZDR vs Confidential for the decision. The current response does not expose a provider-backed versus Reticle-verifiable assurance profile for each model route. A model alias does not inherit confidential-runtime properties without that evidence. See Platform Status.

Get Started

You can integrate Prem API into an application, or you can evaluate Prem API for your organization. The procedure is the same:

Install the SDK

npm install @premai/api-sdk

Create a client with your encryption key

import { createRvencClient } from "@premai/api-sdk";

const client = await createRvencClient({
  apiKey: process.env.PREM_API_KEY,
  clientKEK: process.env.CLIENT_KEK, // You generate this. We never see it.
});

Send a chat request

const response = await client.chat.completions.create({
  model: "deepseek-v4-flash",
  messages: [{ role: "user", content: "Hello, privately." }],
});

For Python, Go, or another language, use the bundled local proxy with a client that supports a custom base URL and the documented routes. Check the Confidential Proxy compatibility limits before rollout.

How It Works

Learn the architecture: the function of each component and the data flow through the system.

Security Model

Examine the trust model: TEEs, attestation, the threat model, and the known limitations.

Developer Experience

Read the integration guide: SDK options, capabilities, code examples, and the API reference.

Platform Status

See the features that are available today, the features that are not ready, and the roadmap.

Agents & Automation

Building an unattended or agentic system? Start here.

Zero Data Retention

The standard mode. It is OpenAI-compatible, it uses trusted partners, and nobody keeps your content.

Go to the Quickstart guide to start immediately.

Get Started

Zero Data Retention

Core Concepts

Handling Errors

What is Prem API

The Problem

What Prem API Does

Client-side encryption

Gateway boundary

Protected processing

The Comparison with a Standard AI API

Who Prem API Is For

Regulated Industries

AI Application Builders

Security-First Organizations

Enterprises with Sensitive IP

What You Can Do Today

Get Started

What to Read Next

How It Works

Security Model

Developer Experience

Platform Status

Agents & Automation

Zero Data Retention

​The Problem

​What Prem API Does

Client-side encryption

Gateway boundary

Protected processing

​The Comparison with a Standard AI API

​Who Prem API Is For

Regulated Industries

AI Application Builders

Security-First Organizations

Enterprises with Sensitive IP

​What You Can Do Today

​Get Started

​What to Read Next

How It Works

Security Model

Developer Experience

Platform Status

Agents & Automation

Zero Data Retention

The Problem

What Prem API Does

The Comparison with a Standard AI API

Who Prem API Is For

What You Can Do Today

Get Started

What to Read Next