Tyko Client - Python SDK

Official Python SDK for Tyko Labs.

Track experiments, manage models, and version datasets with a simple, intuitive API.

Hierarchy

Tyko uses a three-level hierarchy to organize your ML work:

Project → Experiment → Run

• Project: Top-level container for your ML project (e.g., "mnist-classifier")
• Experiment: Groups related runs for comparison (e.g., "hyperparameter-search")
• Run: A single training execution with parameters, metrics, and artifacts

Installation

Install via pip:

pip install tyko

Quick Start

from tyko import TykoClient

client = TykoClient()

# Simplest usage - just project name (uses "default" experiment)
# Environment info (Python version, CPU, GPU, etc.) is auto-captured
with client.start_run(project="my-ml-project") as run:
    run.params["learning_rate"] = 0.001
    run.params["batch_size"] = 32
    # ... your training code ...

# With params at creation time
with client.start_run(
    project="my-ml-project",
    experiment="hyperparameter-search",
    params={"learning_rate": 0.01, "batch_size": 64}
) as run:
    # Params are already set, can add more during the run
    run.params["epochs"] = 100
    # ... your training code ...

Environment Capture

Environment information is automatically captured when you start a run:

• Python version
• Operating system/platform
• CPU count
• RAM size (if psutil is installed)
• GPU count and names (if torch is available)

You can also manually add environment details:

with client.start_run(project="ml-experiments") as run:
    # Add custom environment info
    run.environment["git_commit"] = "abc123"
    run.environment["cuda_version"] = "12.1"

To use the standalone function:

from tyko import capture_environment

env = capture_environment()
print(env)  # {'python_version': '3.12.1', 'platform': 'Linux-...', ...}

Metric Logging

Log metrics during training using run.log():

with client.start_run(project="image-classifier") as run:
    run.params["learning_rate"] = 0.001
    run.params["batch_size"] = 32

    for epoch in range(100):
        train_loss = train_epoch(model, train_data)
        val_loss, val_acc = evaluate(model, val_data)

        # Log multiple metrics at once
        run.log({
            "train/loss": train_loss,
            "eval/loss": val_loss,
            "eval/accuracy": val_acc,
            "epoch": epoch,
        })

Metric Naming Conventions

Use prefixes with slashes for grouped visualization in the dashboard:

This allows the dashboard to organize metrics into logical groups automatically.

Configuration

API Key

Set your API key via environment variable (recommended):

export TYKO_API_KEY="your-api-key"

Or pass it explicitly:

client = TykoClient(api_key="your-api-key")

If the server has anonymous access enabled (ALLOW_ANONYMOUS=true), no API key is needed:

client = TykoClient(api_url="https://your-server.com")

Server URL

For self-hosted deployments, set the server URL:

export TYKO_API_URL="https://your-server.com"

Or:

client = TykoClient(api_url="https://your-server.com")

Development

Prerequisites

• Python 3.11+
• uv package manager

Setup

cd packages/tyko-client
uv sync

Testing

moon run tyko-client:test

Linting

# Check code style
moon run tyko-client:ruff

# Format code
moon run tyko-client:black

# Type checking
moon run tyko-client:mypy

# All three
moon run tyko-client:lint

Building

moon run tyko-client:build