@libertai/confidential-inference

A client for LibertAI models that run inside AMD SEV-SNP virtual machines, where the operator of the machine cannot read the requests being processed. It checks the server's attestation before sending anything.

npm install @libertai/confidential-inference openai

import OpenAI from "openai";
import { connect } from "@libertai/confidential-inference";

const tee = await connect({ model: "qwen3.8-27b-tee" });
const openai = new OpenAI({ apiKey, baseURL: tee.baseURL, fetch: tee.fetch });

const answer = await openai.chat.completions.create({
  model: "qwen3.8-27b-tee",
  messages: [{ role: "user", content: "..." }],
});

API keys come from console.libertai.io. Requests go straight to the enclave, so nothing in between can read them, LibertAI's API included.

Pass tee.fetch along with tee.baseURL: the built-in fetch reaches the same address without checking anything about it.

connect(options)

| Option | | | --- | --- | | model | Use the active deployment for that model. | | itemHash | Pin one deployment and skip discovery. | | tcbFloor | Reject CPUs below a firmware version. Raise it when AMD publishes an advisory. | | signal | AbortSignal. | | publisher, api, scheduler | Override the manifest publisher and the Aleph endpoints. |

Resolves to { baseURL, fetch, itemHash, measurement, sourceCommit }, or throws if the server fails any check, in which case no request was sent.

model is resolved through a manifest signed by LibertAI, naming a deployment message whose hash the client recomputes from its content. The address itself comes from a scheduler that is not trusted: a wrong host just fails attestation.

What gets checked

Before the first byte of a request is written:

• AMD signed the attestation report, through ARK → ASK → VCEK. Only the per-chip VCEK is fetched, and it is self-authenticating.
• The report is bound to the TLS key being served, so a genuine report cannot be replayed in front of someone else's key.
• The launch measurement is one the deployment published, which is what ties the server to a specific model, image and set of serving flags.
• The guest is not debuggable, which would otherwise let the host read its memory.

A measurement identifies which image booted, not what that image does. To check that, rebuild it from source and compare: VERIFYING.md.

Node only

Reading the certificate before any request goes out needs a TLS API that browsers do not expose. A browser would need a local verifying proxy, or a TLS stack in WebAssembly over WebTransport.