Back to UI-TARS-desktop

Introduction

multimodal/websites/tarko/docs/en/guide/get-started/introduction.mdx

0.3.04.7 KB
Original Source

Introduction

Tarko (Tool-augmented Agent Runtime Kernel, Open-source), pronounced /ˈtɑːrkoʊ/, is a tool-call and event-stream driven Agent runtime framework with powerful Context Engineering capabilities.

What is Tarko?

Tarko provides:

Key Features

🔧 Tool Call Engine

Support for multiple Tool Call engines including native Function Call, custom parsers, and models that don't support native Function Tool.

🌊 Event Stream

Built-in event stream processing with native streaming support and standardized Agent Protocol.

🧠 Context Engineering

Advanced context management for long-running agent operations with automatic context optimization.

🔌 Agent Hooks

Extensible hook system for customizing agent behavior at every stage of execution.

Architecture Overview

Tarko is built with a three-layer architecture:

  1. Engineering Layer: CLI, Server, and UI components
  2. Application Layer: Tarko-based Agent implementations
  3. Kernel Layer: Core runtime with Context Engineering and Agent Protocol

Who Uses Tarko?

Tarko powers several production systems:

Documentation Overview

Getting Started

Core Concepts

  • Configuration - Configure agents, models, and runtime settings
  • Model Providers - OpenAI, Anthropic, Volcengine, and custom providers
  • Tools - Create and register custom tools for your agents
  • Tool Call Engine - Native, prompt-engineering, and structured outputs
  • Event Stream - Real-time agent execution monitoring

Advanced Features

UI Integration

Deployment

API Reference

Examples

Quick Navigation

New to Tarko? Start with Quick Start

Building agents? Check out SDK Reference and Tools

Need customization? Explore Agent Hooks and Context Engineering

Deploying to production? See Server Deployment and Agent Snapshot

Building UI? Start with UI Integration