Core Architecture

OpenManus is built on an agent-based architecture with a hierarchical inheritance pattern:

1. BaseAgent (app/agent/base.py): The foundation class that provides:
   • Memory management
   • State transitions
   • Execution loop control
   • Stuck state detection
2. ReActAgent (app/agent/react.py): Extends BaseAgent with:
   • Think-Act cycle pattern
   • Abstract methods for thinking and acting
3. ToolCallAgent (app/agent/toolcall.py): Extends ReActAgent with:
   • Tool/function calling capabilities
   • Tool execution handling
   • Response processing
4. BrowserAgent (app/agent/browser.py): Extends ToolCallAgent with:
   • Browser control capabilities
   • Screenshot handling
   • Browser state management
5. Manus (app/agent/manus.py): The main agent class that extends BrowserAgent with:
   • Additional tools (Python execution, browser use, string replacement)
   • Context-aware thinking process

Main Components

1. Entry Points:
   • main.py: Simple CLI interface to run the Manus agent
   • run_mcp.py: Alternative version using MCP tool
   • run_flow.py: Experimental multi-agent version
2. Tools System (app/tool/):
   • browser_use_tool.py: Web browser automation
   • python_execute.py: Safe Python code execution
   • str_replace_editor.py: Text manipulation
3. LLM Integration (app/llm.py):
   • Handles communication with language models
   • Manages tool calling format
4. Memory and State Management (app/schema.py):
   • Defines message structure
   • Maintains agent state
   • Handles tool calls
5. Configuration (app/config.py):
   • Manages system settings
   • Browser configuration

Execution Flow

When you run python main.py:

1. The script creates a Manus agent instance
2. It prompts the user for input
3. The agent processes the input through its run method
4. The agent executes a series of think-act cycles:
   • Think: Processes the current state and decides what to do
   • Act: Executes the decided actions using tools
5. The agent continues until it reaches a conclusion or hits the maximum step limit

Key Features

1. Browser Automation: Can navigate websites, interact with elements, and extract content
2. Python Code Execution: Can run Python code in a sandboxed environment
3. Planning Capabilities: Uses LLM for planning and reasoning
4. Error Handling: Detects stuck states and has robust error recovery
5. Extensibility: Modular design allows for adding new tools and capabilities

How to Use OpenManus

As seen in the README.md, you can run OpenManus in several ways:

1. Basic usage: python main.py
2. MCP tool version: python run_mcp.py
3. Multi-agent version: python run_flow.py

The agent is designed to handle a wide range of tasks through natural language instructions, leveraging its tools to accomplish complex goals.