zkai

OpenAI-compatible Python SDK for ZKai — private, verifiable AI inference on 0G chain.

ZKai sends your prompt through a Trusted Execution Environment (Intel TDX), encrypts it client-side so even the gateway cannot read it, and anchors a SHA-256 attestation of every inference on the 0G chain. You get back a normal OpenAI-style response plus an on-chain receipt.

Install

pip install zkai

Optional LangChain adapter:

pip install "zkai[langchain]"

Quick start

from zkai import ZKai

client = ZKai(api_key="zkai-...")  # get one at https://zkai-ether-og.vercel.app

response = client.chat.completions.create(
    model="qwen2.5:1.5b",
    messages=[{"role": "user", "content": "Hello!"}],
)

print(response.choices[0].message.content)
print("Attestation hash:", response.attestation_hash)

The response is decrypted locally — the gateway sees only ciphertext. The attestation hash is anchored on 0G mainnet and can be verified independently.

How it works (gateway mode, default)

1. The SDK fetches the enclave's X25519 public key from the gateway.
2. It encrypts your prompt locally with ECDH + ChaCha20-Poly1305.
3. It sends only the ciphertext to the gateway.
4. The gateway routes the opaque blob to a TDX-sealed enclave running the requested model.
5. The enclave decrypts inside sealed memory, runs the model, encrypts the response, and emits a SHA-256 attestation hash.
6. The attestation lands on the on-chain AttestationRegistry contract.
7. The SDK decrypts the response locally and returns it to you.

End-to-end, the gateway never sees plaintext.

Configuration

ZKai(
    api_key="zkai-...",                             # issued from the dashboard
    base_url="https://zkai-ether-og.vercel.app",    # default
    encrypted=True,                                  # default; set False to use the legacy plaintext path
    provider_endpoint=None,                          # set to bypass gateway and hit a provider directly
    skip_attestation=False,                          # only for development; do not disable in prod
)

OpenAI compatibility

The response object mirrors OpenAI's chat completion shape:

response.id
response.model
response.choices[0].message.content
response.choices[0].finish_reason
response.usage.prompt_tokens
response.usage.completion_tokens

# ZKai-specific:
response.attestation_hash  # on-chain commitment

Drop-in replacement for most OpenAI SDK call sites — just swap the client class.

On-chain components (0G mainnet, chain ID 16661)

Verify any attestation hash on the 0G explorer.

Running your own provider

See the companion zkai-cli package:

pip install zkai-cli
zkai init && zkai start && zkai register

Links

• Dashboard — https://zkai-ether-og.vercel.app
• Repository — https://github.com/skyyycodes/zkai-eth
• Provider CLI — https://pypi.org/project/zkai-cli/

License

MIT