UI-TARS Desktop and Agent Stack

The Agent TARS Web UI provides a feature-complete interface for managing sessions and viewing MCP Tool Results. Key capabilities include:

• Native Streaming: Real-time streaming support (available since Agent TARS Beta).
• GUI Grounding Transition: Supports GUI Grounding with real-time cursor animations to visualize agent actions.
• Share and Reply: Supports saving replays locally by default. Users can also configure a share.provider to upload replays to a custom server.
• Dark Mode Support: Supports both Dark and Light modes, with Dark mode enabled by default. Theme preferences are saved locally.

The @tarko/interface package provides high-level interfaces for the Tarko ecosystem. It specifically includes the Tarko Agent Server and the Tarko Agent CLI.

If you require lower-level access to the Agent Interface, you should use the @tarko/agent-interface package instead.

The @ui-tars/operator-adb package is an Android SDK operator for UI-TARS that leverages ADB (Android Debug Bridge) to interact with devices.

Key capabilities include:

• Screen capture: Capturing the device screen via ADB.
• Interaction simulation: Simulating touch and swipe gestures.
• Chinese input: Support for Chinese character input via ADBKeyBoard.

Agent TARS is an open-source multimodal AI agent designed for human-like task completion. It leverages rich multimodal capabilities including GUI Agent (visual grounding) and Vision.

It features built-in integration with several real-world tools:

• Search: Web search capabilities.
• Browser: Web browsing and interaction.
• File: File system operations.
• Command: Terminal/command execution.

Agent TARS provides two primary ways to interact with it:

1. CLI: Best for quickly experiencing the agent's capabilities.
2. SDK: Designed for developers who want to build their own custom agents using the Agent TARS framework.

TARS is a multimodal AI Agent stack consisting of two primary projects:

1. Agent TARS: A general multimodal AI Agent stack designed to bring GUI Agent and Vision capabilities to your terminal, computer, browser, and products. It is primarily used via a CLI or Web UI and integrates with MCP (Model Context Protocol) tools to achieve human-like task completion.

2. UI-TARS Desktop: A desktop application providing a native GUI Agent based on the UI-TARS model. It supports both local and remote computer and browser operators.

Tarko (Tool-augmented Agent Runtime Kernel) is an open-source, tool-call and event-stream driven agent runtime framework. It is designed to support long-running agent operations through advanced context management and extensible execution patterns.

Core Capabilities

• Context Engineering: Advanced management and automatic optimization of context for long-running operations.
• Tool Call Engine: Support for multiple LLM providers using native function calling, custom parsers, or prompt-engineering for models without native tool support.
• Event Stream Processing: Native streaming support and adherence to the standardized Agent Protocol for real-time monitoring.
• Agent Hooks: An extensible system to customize agent behavior at various lifecycle stages.
• Agent Protocol: Standardized format definitions for managing the agent lifecycle.

UI-TARS Desktop is a GUI Agent application powered by the UI-TARS Vision-Language Model. It enables users to control their computers (Windows and MacOS) using natural language instructions. The agent performs visual recognition via screenshots and executes precise mouse and keyboard controls to interact with the operating system and applications.

Key features include:

• Natural Language Control: Interact with your computer using plain text.
• Visual Recognition: Uses screenshots to understand the current state of the screen.
• Precise Control: Executes mouse clicks, movements, and keyboard inputs.
• Cross-platform: Supports both Windows and MacOS.
• Privacy: Supports fully local processing for secure automation.

The @agent-infra/browser library provides the following capabilities:

• Browser Detection: Automatically detects installed browsers across different platforms.
• Remote Browser Support: Ability to connect to remote browser instances.
• Type Safety: Full TypeScript support for reliable development.