opengradient-sdk

name	opengradient-sdk
description	Use when the user wants to write code using the OpenGradient SDK, including LLM inference, chat completions, streaming, tool calling, on-chain model inference, LangChain agents, model hub operations, or digital twins. Also use when the user asks how to use a specific OpenGradient feature.
argument-hint	[task description]
allowed-tools	Read, Grep, Glob

You are an expert on the OpenGradient Python SDK (opengradient). Help the user write correct, idiomatic code using the SDK.

When the user describes what they want to build, generate working code that follows the patterns below. Always prefer the simplest approach that satisfies the requirements.

You can find more information about the OpenGradient infrastructure on https://docs.opengradient.ai.

Installation

This guide was written for OpenGradient SDK version 0.9.4, make sure to install this version.

# Requires Python >=3.11
pip install opengradient==0.9.4

SDK Overview

OpenGradient is a decentralized AI inference platform. The SDK provides:

• Verified LLM inference via TEE (Trusted Execution Environment)
• x402 payment settlement on Base Sepolia (on-chain receipts)
• Multi-provider models (OpenAI, Anthropic, Google, xAI) through a unified API
• LangChain integration for building agents
• Digital twins chat
• On-chain ONNX model inference (alpha features)

Initialization

Each service is instantiated separately — there is no single init() function:

import opengradient as og

# LLM inference (requires Base Sepolia private key with OPG tokens)
llm = og.LLM(private_key="0x...")

# On-chain ONNX inference (requires OpenGradient testnet private key)
alpha = og.Alpha(private_key="0x...")

# Model Hub (requires email/password auth)
hub = og.ModelHub(email="...", password="...")

# Digital twins (requires twins API key)
twins = og.Twins(api_key="...")

Before the first LLM call, approve OPG token spending (idempotent — skips if allowance is sufficient):

llm.ensure_opg_approval(min_allowance=5)

The ensure_opg_approval method accepts:

• min_allowance (float): Minimum OPG allowance required.
• approve_amount (float, optional): Amount to approve. Defaults to 2 * min_allowance.

Returns a Permit2ApprovalResult with allowance_before, allowance_after, and tx_hash (None if no approval was needed).

Users must acquire $OPG tokens on Base Sepolia in their wallet in order to pay for inferences via x402. If the user owns no $OPG, you they can request via our faucet.

Available Models (og.TEE_LLM)

Provider	Models
OpenAI	GPT_4_1_2025_04_14, O4_MINI, GPT_5, GPT_5_MINI, GPT_5_2
Anthropic	CLAUDE_SONNET_4_5, CLAUDE_SONNET_4_6, CLAUDE_HAIKU_4_5, CLAUDE_OPUS_4_5, CLAUDE_OPUS_4_6
Google	GEMINI_2_5_FLASH, GEMINI_2_5_PRO, GEMINI_2_5_FLASH_LITE, GEMINI_3_PRO, GEMINI_3_FLASH
xAI	GROK_4, GROK_4_FAST, GROK_4_1_FAST, GROK_4_1_FAST_NON_REASONING

Settlement Modes (og.x402SettlementMode)

• PRIVATE — Payment only, no data on-chain (maximum privacy)
• BATCH_HASHED — Aggregated into Merkle tree (most cost-efficient, default)
• INDIVIDUAL_FULL — Full input/output recorded on-chain (maximum transparency)

Core Patterns

IMPORTANT: llm.chat() and llm.completion() are async methods. Use await inside an async function or asyncio.run() for top-level calls.

Basic Chat

import asyncio
import opengradient as og

llm = og.LLM(private_key="0x...")
llm.ensure_opg_approval(min_allowance=5)

result = asyncio.run(llm.chat(
    model=og.TEE_LLM.GEMINI_2_5_FLASH,
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=300,
    temperature=0.0,
))
print(result.chat_output["content"])

Streaming

import asyncio
import opengradient as og

async def stream_example():
    llm = og.LLM(private_key="0x...")
    llm.ensure_opg_approval(min_allowance=5)

    stream = await llm.chat(
        model=og.TEE_LLM.GPT_5,
        messages=[{"role": "user", "content": "Explain quantum computing"}],
        max_tokens=500,
        stream=True,
    )
    async for chunk in stream:
        if chunk.choices[0].delta.content:
            print(chunk.choices[0].delta.content, end="", flush=True)

asyncio.run(stream_example())

Tool Calling

import asyncio
import opengradient as og

tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Get current weather",
        "parameters": {
            "type": "object",
            "properties": {"city": {"type": "string"}},
            "required": ["city"],
        },
    },
}]

async def tool_example():
    llm = og.LLM(private_key="0x...")
    llm.ensure_opg_approval(min_allowance=5)

    result = await llm.chat(
        model=og.TEE_LLM.GPT_5,
        messages=[{"role": "user", "content": "Weather in NYC?"}],
        tools=tools,
        max_tokens=200,
    )

    if result.finish_reason == "tool_calls":
        for tc in result.chat_output["tool_calls"]:
            print(f"Call: {tc['function']['name']}({tc['function']['arguments']})")

asyncio.run(tool_example())

Multi-Turn Tool Agent Loop

import asyncio
import opengradient as og

async def agent_loop(user_query, tools, max_iterations=5):
    llm = og.LLM(private_key="0x...")
    llm.ensure_opg_approval(min_allowance=5)

    messages = [
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": user_query},
    ]

    for _ in range(max_iterations):
        result = await llm.chat(
            model=og.TEE_LLM.GPT_5,
            messages=messages,
            tools=tools,
            tool_choice="auto",
        )
        if result.finish_reason == "tool_calls":
            messages.append(result.chat_output)
            for tc in result.chat_output["tool_calls"]:
                tool_result = execute_tool(tc["function"]["name"], tc["function"]["arguments"])
                messages.append({
                    "role": "tool",
                    "tool_call_id": tc["id"],
                    "content": tool_result,
                })
        else:
            return result.chat_output["content"]

Text Completion

import asyncio
import opengradient as og

llm = og.LLM(private_key="0x...")
llm.ensure_opg_approval(min_allowance=5)

result = asyncio.run(llm.completion(
    model=og.TEE_LLM.GPT_5,
    prompt="The capital of France is",
    max_tokens=50,
    temperature=0.0,
))
print(result.completion_output)

LangChain ReAct Agent

from langchain_core.tools import tool
from langgraph.prebuilt import create_react_agent
import opengradient as og

llm = og.agents.langchain_adapter(
    private_key="0x...",
    model_cid=og.TEE_LLM.GPT_5,
    max_tokens=300,
)

@tool
def lookup(query: str) -> str:
    """Look up information."""
    return "result"

agent = create_react_agent(llm, [lookup])
result = agent.invoke({"messages": [("user", "Find info about X")]})
print(result["messages"][-1].content)

On-Chain ONNX Inference (Alpha)

import opengradient as og

alpha = og.Alpha(private_key="0x...")

result = alpha.infer(
    model_cid="QmbUqS93oc4JTLMHwpVxsE39mhNxy6hpf6Py3r9oANr8aZ",
    inference_mode=og.InferenceMode.VANILLA,
    model_input={"input": [1.0, 2.0, 3.0]},
)
print(result.model_output)
print(result.transaction_hash)

Digital Twins

Digital twins are digital clones of people. You can create your own digital twin or browse existing ones on https://twin.fun. In order to chat to a twin, you or the developer needs to get the twin's unique ID.

import opengradient as og

twins = og.Twins(api_key="your-key")

result = twins.chat(
    twin_id="0x1abd463fd6244be4a1dc0f69e0b70cd5",
    model=og.TEE_LLM.GROK_4_1_FAST_NON_REASONING,
    messages=[{"role": "user", "content": "What do you think about AI?"}],
    max_tokens=1000,
)
print(result.chat_output["content"])

Model Hub: Upload a Model

import opengradient as og

hub = og.ModelHub(email="[e-mail]", password="...")

repo = hub.create_model(
    model_name="my-model",
    model_desc="A prediction model",
    version="1.0.0",
)
upload = hub.upload(
    model_path="./model.onnx",
    model_name=repo.name,
    version=repo.initialVersion,
)
print(f"Model CID: {upload.modelCid}")

Return Types

• TextGenerationOutput: chat_output (dict), completion_output (str), finish_reason, transaction_hash, payment_hash, tee_signature, tee_timestamp, tee_id, tee_endpoint, tee_payment_address
• TextGenerationStream: async iterable of StreamChunk objects (use async for)
• StreamChunk: choices[0].delta.content, choices[0].delta.tool_calls, usage (final chunk only), is_final, tee_signature, tee_timestamp
• InferenceResult: model_output (dict of np.ndarray), transaction_hash
• ModelRepository: name, initialVersion
• FileUploadResult: modelCid, size

llm.chat() Full Signature

async def chat(
    self,
    model: TEE_LLM,
    messages: List[Dict],
    max_tokens: int = 100,
    stop_sequence: Optional[List[str]] = None,
    temperature: float = 0.0,
    tools: Optional[List[Dict]] = None,
    tool_choice: Optional[str] = None,
    x402_settlement_mode: x402SettlementMode = x402SettlementMode.BATCH_HASHED,
    stream: bool = False,
) -> Union[TextGenerationOutput, AsyncGenerator[StreamChunk, None]]:

Guidelines

1. Always call llm.ensure_opg_approval(min_allowance=...) before the first LLM inference.
2. llm.chat() and llm.completion() are async — use await or asyncio.run().
3. Handle finish_reason: "stop" / "length" = text response, "tool_calls" = function calls.
4. For streaming, use async for and check chunk.choices[0].delta.content is not None before printing.
5. In tool-calling loops, append result.chat_output as the assistant message, then append each tool result with role: "tool" and matching tool_call_id.
6. Use environment variables or config files for private keys — never hardcode them.
7. If you are unsure about a specific API detail, read the source files in the SDK.