tensorzero 2026.5.0

The Python client for TensorZero

TensorZero Python Client

Website · Docs · Twitter · Slack · Discord

Quick Start (5min) · Deployment Guide · API Reference · Configuration Reference

The tensorzero package provides a Python client for the TensorZero Gateway. This client allows you to easily make inference requests and assign feedback to them via the gateway.

See our API Reference for more information.

Installation

pip install tensorzero

Basic Usage

Initialization

The TensorZero client offers synchronous (TensorZeroGateway) and asynchronous (AsyncTensorZeroGateway) variants. Additionally, the client can launch an embedded (in-memory) gateway (build_embedded) or connect to an external HTTP gateway (build_http) - both of these methods return a gateway instance.

By default, the asynchronous client returns a Future when you call build_http or build_embedded, so you must await it. If you prefer to avoid the await, you can set async_setup=False to initialize the client in a blocking way.

Synchronous HTTP Gateway

from tensorzero import TensorZeroGateway

with TensorZeroGateway.build_http(gateway_url="http://localhost:3000") as client:
    # ...

Asynchronous HTTP Gateway

import asyncio

from tensorzero import AsyncTensorZeroGateway


async def run():
    async with await AsyncTensorZeroGateway.build_http(
        gateway_url="http://localhost:3000",
        # async_setup=False  # optional: skip the `await` and run `build_http` synchronously (blocking)
    ) as client:
        # ...


if __name__ == "__main__":
    asyncio.run(run())

Synchronous Embedded Gateway

from tensorzero import TensorZeroGateway

with TensorZeroGateway.build_embedded(
    config_file="/path/to/tensorzero.toml",
    clickhouse_url="http://chuser:chpassword@localhost:8123/tensorzero"
) as client:
    # ...

Asynchronous Embedded Gateway

import asyncio

from tensorzero import AsyncTensorZeroGateway


async def run():
    async with await AsyncTensorZeroGateway.build_embedded(
        config_file="/path/to/tensorzero.toml",
        clickhouse_url="http://chuser:chpassword@localhost:8123/tensorzero"
        # async_setup=False  # optional: skip the `await` and run `build_embedded` synchronously (blocking)
    ) as client:
        # ...


if __name__ == "__main__":
    asyncio.run(run())

Inference

Non-Streaming Inference with Synchronous Client

with TensorZeroGateway.build_http(gateway_url="http://localhost:3000") as client:
    response = client.inference(
        model_name="openai::gpt-4o-mini",
        input={
            "messages": [
                {"role": "user", "content": "What is the capital of Japan?"},
            ],
        },
    )

    print(response)

Non-Streaming Inference with Asynchronous Client

async with await AsyncTensorZeroGateway.build_http(gateway_url="http://localhost:3000") as client:
    response = await client.inference(
        model_name="openai::gpt-4o-mini",
        input={
            "messages": [
                {"role": "user", "content": "What is the capital of Japan?"},
            ],
        },
    )

    print(response)

Streaming Inference with Synchronous Client

with TensorZeroGateway.build_http(gateway_url="http://localhost:3000") as client:
    stream = client.inference(
        model_name="openai::gpt-4o-mini",
        input={
            "messages": [
                {"role": "user", "content": "What is the capital of Japan?"},
            ],
        },
        stream=True,
    )

    for chunk in stream:
        print(chunk)

Streaming Inference with Asynchronous Client

async with await AsyncTensorZeroGateway.build_http(gateway_url="http://localhost:3000") as client:
    stream = await client.inference(
        model_name="openai::gpt-4o-mini",
        input={
            "messages": [{"role": "user", "content": "What is the capital of Japan?"}],
        },
        stream=True,
    )

    async for chunk in stream:
        print(chunk)

Feedback

Synchronous

with TensorZeroGateway.build_http(gateway_url="http://localhost:3000") as client:
    response = client.feedback(
        metric_name="thumbs_up",
        inference_id="00000000-0000-0000-0000-000000000000",
        value=True,  # 👍
    )

    print(response)

Asynchronous

async with await AsyncTensorZeroGateway.build_http(gateway_url="http://localhost:3000") as client:
    response = await client.feedback(
        metric_name="thumbs_up",
        inference_id="00000000-0000-0000-0000-000000000000",
        value=True,  # 👍
    )

    print(response)

Stubtest

Run uv run stubtest tensorzero.tensorzero to confirm that the stub tensorzero.pyi file matches PyO3 compilation.