Skip to content

AxLabs/venice-e2ee

 
 

Repository files navigation

@axlabs/venice-e2ee

npm License: GPL-3.0

End-to-end encryption for Venice AI's TEE-backed inference, maintained by AxLabs. Prompts are encrypted client-side and only decrypted inside Intel TDX Trusted Execution Environments — Venice never sees plaintext.

Note: This library uses standard cryptographic primitives (ECDH, HKDF, AES-256-GCM) via audited implementations (@noble/secp256k1, Web Crypto API). It implements no custom cryptography — it packages Venice's E2EE protocol into a reusable library. Review SECURITY.md for the trust model and known limitations before relying on it.

Protocol: ECDH (secp256k1) key exchange → HKDF-SHA256 key derivation → AES-256-GCM encryption

Install

npm install @axlabs/venice-e2ee

Or use the browser bundle directly:

<script type="module">
  import { createVeniceE2EE } from './venice-e2ee.browser.js';
</script>

Usage

import { createVeniceE2EE } from '@axlabs/venice-e2ee';

const e2ee = createVeniceE2EE({ apiKey: 'your-venice-api-key' });

// Create session (fetches TEE attestation, verifies quote, ECDH key exchange)
const session = await e2ee.createSession('e2ee-qwen3-5-122b-a10b');

// Inspect attestation result
console.log(session.attestation);
// { nonceVerified: true, signingKeyBound: true, debugMode: false, serverTdxValid: true, errors: [] }

// Encrypt messages
const { encryptedMessages, headers, veniceParameters } = await e2ee.encrypt(
  [{ role: 'user', content: 'Hello from the encrypted side' }],
  session
);

// Send to Venice API
const response = await fetch('https://api.venice.ai/api/v1/chat/completions', {
  method: 'POST',
  headers: { 'Content-Type': 'application/json', Authorization: `Bearer ${apiKey}`, ...headers },
  body: JSON.stringify({
    model: 'e2ee-qwen3-5-122b-a10b',
    messages: encryptedMessages,
    stream: true,
    venice_parameters: veniceParameters,
  }),
});

// Decrypt streaming response
for await (const chunk of e2ee.decryptStream(response.body, session)) {
  process.stdout.write(chunk);
}

API

createVeniceE2EE(options)

Creates an E2EE instance with session caching and attestation verification.

Option Type Default Description
apiKey string required Venice API key
baseUrl string https://api.venice.ai API base URL
sessionTTL number 1800000 (30 min) Session cache TTL in ms
verifyAttestation boolean true Verify TEE attestation on session creation
dcapVerifier DcapVerifier Optional full DCAP verifier (see below)

Returns an object with:

  • createSession(modelId) — Generates ephemeral keypair, fetches TEE attestation, verifies the TDX quote, derives AES key. Returns an E2EESession with attestation results. Throws if verification fails. Sessions are cached per model with a 30-minute TTL.
  • encrypt(messages, session) — Encrypts an array of {role, content} messages. Returns { encryptedMessages, headers, veniceParameters }.
  • decryptChunk(hexChunk, session) — Decrypts a single response chunk (hex-encoded ciphertext with embedded server ephemeral key).
  • decryptStream(body, session) — Async generator that parses an SSE stream and yields decrypted text chunks. Handles per-chunk ephemeral keys, plaintext passthrough, and [DONE] sentinel.
  • clearSession() — Zeroizes the private key and clears the cached session.

Attestation verification

Every createSession call fetches a TDX attestation quote from Venice and verifies it client-side:

  1. Nonce binding — confirms the client nonce appears in REPORTDATA (raw or SHA-256)
  2. Signing key binding — confirms the signing key's Ethereum address matches REPORTDATA
  3. Debug mode rejection — rejects TEEs running in debug mode
  4. Server cross-check — flags inconsistencies with Venice's own verification results

If any check fails, createSession throws with a descriptive error. The attestation result is available on session.attestation.

To disable verification (not recommended):

const e2ee = createVeniceE2EE({ apiKey, verifyAttestation: false });

Full DCAP verification (optional)

For full TDX DCAP verification (PCK cert chain, quote signatures, TCB evaluation), install the optional peer dependency and inject the verifier:

npm install @phala/dcap-qvl
import { createVeniceE2EE } from '@axlabs/venice-e2ee';
import { createDcapVerifier } from '@axlabs/venice-e2ee/dcap';

const e2ee = createVeniceE2EE({
  apiKey: 'your-venice-api-key',
  dcapVerifier: createDcapVerifier(), // uses Phala PCCS by default
});

This adds ~500KB to browser bundles. For most use cases, the default binding checks + server cross-check are sufficient.

isE2EEModel(modelId)

Returns true if the model ID starts with e2ee-.

Low-level exports

For custom integrations, the individual crypto and attestation primitives are also exported:

import {
  generateKeypair,      // secp256k1 ephemeral keypair
  deriveAESKey,         // ECDH shared secret → HKDF → AES-256-GCM key
  encryptMessage,       // AES-GCM encrypt → hex(pubkey + nonce + ciphertext)
  decryptChunk,         // per-chunk ECDH + AES-GCM decrypt
  decryptSSEStream,     // SSE parser + decryption async generator
  verifyAttestation,    // run attestation checks on a raw response
  deriveEthAddress,     // secp256k1 pubkey → Ethereum address
  toHex,
  fromHex,
} from '@axlabs/venice-e2ee';

How it works

Client                              Venice TEE (Intel TDX)
  |                                        |
  |── GET /tee/attestation?model=&nonce= ─>|
  |<── { signing_key, intel_quote, ... } ──|
  |                                        |
  |  Parse TDX quote                       |
  |  Verify nonce in REPORTDATA            |
  |  Verify signing key address binding    |
  |  Reject debug mode                     |
  |                                        |
  |  generateKeypair()                     |
  |  deriveAESKey(clientPriv, teePub)      |
  |  encryptMessage(aesKey, msg)           |
  |                                        |
  |── POST /chat/completions  ────────────>|
  |   X-Venice-TEE-Client-Pub-Key: ...     |
  |   X-Venice-TEE-Model-Pub-Key: ...      |
  |   { messages: [encrypted] }            |
  |                                        |
  |<── SSE stream (per-chunk encryption) ──|
  |    each chunk: hex(ephemeralPub +       |
  |                     nonce + ciphertext) |
  |                                        |
  |  decryptChunk(clientPriv, chunk)        |
  |  → ECDH(clientPriv, chunkEphPub)       |
  |  → HKDF → AES-GCM decrypt             |

Each response chunk uses a fresh server ephemeral key, so every chunk requires its own ECDH key derivation.

Security

What's verified:

  • Signing key is cryptographically bound to the TEE via TDX REPORTDATA
  • Client nonce prevents replay attacks
  • Debug-mode TEEs are rejected
  • ECDH intermediates are zeroized after key derivation
  • Private keys are zeroized on session clear/replacement

What's not verified client-side (by default):

  • TDX quote signature chain (available via optional DCAP verifier)
  • NVIDIA GPU attestation
  • TEE code measurements (Venice doesn't publish expected values yet)

For the full trust model and known limitations, see SECURITY.md. For the verification design and module layout, see ARCHITECTURE.md.

Development

This repo pins its Node version in .nvmrc. If you use nvm, run nvm use first:

nvm use               # match the pinned Node version
npm install
npm test              # unit + integration tests
npm run build         # TypeScript → dist/
npm run build:browser # single-file ESM bundle

Set VENICE_API_KEY in .env to run integration tests against the live API.

See DEVELOPMENT.md for CI details and the npm release process.

Acknowledgments

License

GPL-3.0 — see LICENSE

Copyright © AxLabs.

About

Venice AI end-to-end encryption library — ECDH secp256k1, AES-256-GCM, TEE attestation verification

Resources

License

Security policy

Stars

Watchers

Forks

Packages

 
 
 

Contributors

Languages

  • TypeScript 94.3%
  • Shell 5.7%