Aworld

AWorld is a multi - agent system framework aiming to bridge the gap between theoretical MAS capabilities and practical applications, providing a full - set solution from single - agent to multi - agent collaboration/competition. The project supports scenarios such as browser/mobile operations and GAIA benchmark testing, adopts a client - server architecture, integrates a rich toolchain, and includes performance evaluation and training functions.

Artificial intelligence chatbots Developer tools #Agent framework #Multi - agent #Tool integration #Environment simulation .Python

rating : 2 points

downloads : 11

update time : 2025-04-29

What is the MCP server?

The MCP server is a core service component in the AWorld framework, providing standardized tool call interfaces and environment interaction capabilities for AI agents. It's like an intelligent toolbox that allows AI agents to use various functions and services safely and efficiently.

How to use the MCP server?

Developers only need to specify the name of the MCP server to be used in the agent configuration, and the AI agent can automatically obtain the corresponding capabilities. Users don't need to care about the underlying implementation details.

Applicable scenarios

It is suitable for scenarios where AI agents need to perform complex tasks, use external tools, or interact with other systems, such as data analysis, web operations, API calls, etc.

Main features

Unified service interfaceProvide a standardized call method for various tools and services, simplifying the development of AI agents

Security sandboxAll tool calls are executed in a controlled environment to ensure system security

Automatic routingIntelligently route requests to the most suitable service instance to optimize performance

Advantages and limitations

Advantages

Ready to use without complex configuration

Support hot plugging and dynamically add new services

Provide a unified monitoring and management interface

Limitations

Service instances need to be pre - deployed

Some advanced functions require professional configuration

Performance depends on the quality of underlying services

How to use

Configure the MCP server

Add the names of the required MCP servers in the agent configuration file

Call the service

The AI agent will automatically obtain the corresponding service capabilities and can be used directly in the conversation

Usage examples

Map service queryUse the map MCP service to query location information

Weather queryObtain real - time weather data through the weather service

Frequently Asked Questions

How to view available MCP services?

What should I do if the call fails?

Related resources

MCP service development guide

How to develop custom MCP services

Example configuration file

Example configuration of the MCP server

🚀 AWorld: Unified Agent Training and Evaluation Platform

AWorld is a unified framework for training and evaluating agents designed for computer and mobile phone usage tasks. It aims to offer a powerful and user - friendly environment for agent developers to test, train, and optimize agent performance in various real - world scenarios.

🚀 Quick Start

📦 Installation

pip install aworld

💻 Usage Examples

Basic Usage

from aworld import World, Agent

# Initialize the world and the agent
world = World()
agent = Agent()

# Submit a task
task_id = world.submit_task("your_task.yaml")
result = agent.execute(task_id)

# Print the result
print(result)

✨ Features

Unified Task Definition

AWorld provides a unified task definition interface, supporting custom tasks and public benchmarks (such as GAIA, WebArena, etc.). Users can quickly define new tasks by writing YAML configuration files.

Stable Execution Environment

The platform offers an efficient and stable execution environment, with built - in debugging tools and monitoring systems to ensure the smooth completion of agent tasks.

Detailed Task Reports

AWorld automatically generates detailed test reports, including the following key metrics:

Efficiency: The number of steps required to complete the task.
Success Rate: The task completion rate.
Cost: The number of tokens consumed.
Time: The duration of task execution.

Agent Training Support

In addition to the evaluation function, AWorld also provides an agent training interface. Users can continuously optimize agent models based on actual task data and manage different versions of agents through a version control system.

🔧 Technical Details

Core Components

Actor

The agent in AWorld is a component with rich functions, capable of performing various tasks and collaborating with other agents. The following are the main attributes of the agent:

Property	Details
`id`	String, the unique identifier of the agent.
`name`	String, the name of the agent.
`model_name`	String, the name of the large language model (LLM).
`_llm`	Object, an LLM instance initialized based on `model_name`.
`conf`	BaseModel, configuration information inherited from pydantic's BaseModel.
`trajectory`	Object, a memory structure for maintaining the agent's context.
`tool_names`	List, a list of tools that the agent can use.
`mcp_servers`	List, a list of MCP servers that the agent can access.
`handoffs`	List, a list of agents to which tasks can be delegated.
`finished`	Boolean, a flag indicating whether the agent has completed the task.

Environment/World Model

AWorld provides a rich toolset and multiple world models to simulate various real - world scenarios:

Tool	Description
`MCP Server`	Integrates multiple MCP servers, providing powerful tool support.
`Browser`	Controls the web browser to perform functions such as navigation and form filling.
`Android`	Simulates Android devices for mobile application testing and automation.
`Shell`	Executes shell commands to manage the file system and system interactions.
`Code`	Runs code snippets, supporting multiple programming languages.
`Search`	Performs web searches and returns structured results.
`Document`	Handles file operations, including reading, writing, and directory management.

📚 Documentation

Project Characteristics

Open - Source and Scalability

AWorld is an open - source project, providing a complete code repository and detailed documentation. It supports function extension through a plugin mechanism, allowing developers to add custom tools and services.

Multi - Modal Support

The platform not only supports text interaction but also integrates multiple third - party services (such as network requests, image recognition, etc.) to provide multi - modal capabilities for agents.