ObelyZK SDK

Unified SDKs for verifiable ML inference on Starknet. Prove any model, verify on-chain in 1 transaction with full OODS + Merkle + FRI + PoW (trustless) verification.

Packages

Package	Language	Install	Description
@obelyzk/sdk	TypeScript	npm install @obelyzk/sdk	Full-featured prover client with async jobs
obelyzk	Python	pip install obelyzk	Pythonic API with sync and async support
@obelyzk/cli	CLI	npm install -g @obelyzk/cli	Command-line proving and submission
@obelyzk/mcp-server	MCP	npm install @obelyzk/mcp-server	Claude AI tool integration (40+ tools)

API Comparison

Feature	TypeScript	Python	CLI
Prove a model	client.prove({ model, input })	client.prove(model, input)	obelysk prove --model --input
List models	client.getModels()	client.models()	obelysk models
Attestation	client.attest({ model, input })	client.attest(model, input)	obelysk prove --on-chain
Async support	Native Promises	AsyncObelyzkClient	Background jobs
Job polling	client.getJob(id)	client.job(id)	obelysk status --job
Config	Constructor options	Constructor args	obelysk config

Quick Start

TypeScript

import { createProverClient } from "@obelyzk/sdk";

const client = createProverClient();
const result = await client.prove({
  model: "smollm2-135m",
  input: [1.0, 2.0, 3.0],
  onChain: true,
});

console.log("Proof TX:", result.txHash);
console.log("Verified:", result.verified);

See the full TypeScript SDK README for API reference and examples.

Python

from obelyzk import ObelyzkClient

client = ObelyzkClient()
result = client.prove(
    model="smollm2-135m",
    input=[1.0, 2.0, 3.0],
    on_chain=True,
)

print(f"Proof TX: {result.tx_hash}")
print(f"Verified: {result.verified}")

See the full Python SDK README for API reference and async usage.

CLI

# Install
npm install -g @obelyzk/cli

# Prove and verify on-chain
obelysk prove --model smollm2-135m --input "Hello world" --on-chain

# List available models
obelysk models

See the full CLI README for all commands and flags.

No SDK (curl)

If you prefer raw HTTP, the hosted API works with plain curl. No SDK installation needed.

# Health check (no auth)
curl https://api.bitsage.network/health

# Prove a model
curl -X POST https://api.bitsage.network/api/v1/infer \
  -H "Authorization: Bearer $OBELYSK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model_id":"smollm2-135m","input":[1.0,2.0,3.0],"gpu":true}'

# Check on-chain verification
starkli call 0x1c208a5fe731c0d03b098b524f274c537587ea1d43d903838cc4a2bf90c40c7 \
  get_recursive_verification_count 0x7214ee0e9c30e3e6748651d42f941c4b875a5b0a549223f92471a58585c980

Auth: Authorization: Bearer <API_KEY> header on every request. Set OBELYSK_API_KEY in your environment.

How It Works

1. Your SDK call hits the hosted GPU prover at https://api.bitsage.network
2. The prover executes the model over the M31 field and generates a GKR sumcheck proof
3. A recursive STARK compresses the proof to ~942 felts (constant size, 49x compression)
4. The proof is verified on Starknet Sepolia in a single transaction using full OODS + Merkle + FRI + PoW (trustless)

On-Chain Verification

All proofs are verified by the ObelyZK Recursive Verifier contract on Starknet Sepolia:

• Contract: 0x1c208a5fe731c0d03b098b524f274c537587ea1d43d903838cc4a2bf90c40c7
• Verification: Full OODS + Merkle + FRI + PoW (trustless)
• Network: Starknet Sepolia
• Felts: ~942 per proof
• Compression: 49x vs raw GKR data
• Cost: ~$0.02 per verification

Supported Models

Model	Params	Prove Time (GPU)	Recursive Felts
SmolLM2-135M	135M	~102s	942
Qwen2-0.5B	500M	~45s	~900
Phi-3-mini	3.8B	~180s	~950
Custom HuggingFace	Any LLaMA/Qwen/Phi	Varies	~950

Environment Variables

Variable	Description	Required
OBELYSK_API_KEY	API key for hosted prover	For hosted
OBELYSK_PROVER_URL	Custom prover URL	For self-hosted
STARKNET_ACCOUNT	Starknet account address	For on-chain
STARKNET_PRIVATE_KEY	Starknet private key	For on-chain

Configuration

All SDKs accept a prover URL. Default is the hosted service:

// Use hosted prover (default)
const client = createProverClient();

// Use your own GPU prover
const client = createProverClient({ url: "http://your-gpu:8080" });

Links

• Main Documentation
• Getting Started Guide
• On-Chain Verification
• npm: @obelyzk/sdk
• PyPI: obelyzk