OpenManus

OpenManus is an open-source framework for building general AI agents, created by contributors from the MetaGPT community as an open alternative to Manus. You run the agent from a terminal, type an idea, and it plans and executes the work using tools such as browser automation, code execution, and web crawling.

Three entry points are provided: main.py for the general agent, runmcp.py for the MCP tool version, and runflow.py for a multi-agent flow that is marked unstable. The flow can include a DataAnalysis agent for analysis and visualisation tasks. LLM access is configured in a config.toml file with a global model and an optional separate vision model, and works with any OpenAI-compatible endpoint. Python 3.12 is required.

Browser automation runs Browser Use CLI 3.0 as a default MCP server through uvx, which keeps Browser Use and its dependencies isolated from the OpenManus environment. Local mode attaches to Chrome or Chromium without an API key; a Browser Use Cloud key enables an isolated remote browser, and existing browsers can be selected via BUCDPURL, BUCDPWS, or BU_NAME. BrowserGym tasks still need a Playwright browser.

Features

• General agent: a single agent that plans and executes tasks from a natural-language prompt
• Browser tools: the Browser Use skill plus native browserexec and browserscreenshot tools
• MCP mode: run the agent with tools served over the Model Context Protocol
• Multi-agent flow: an experimental orchestration of several agents, including a DataAnalysis agent
• Model configuration: global and vision-specific LLM settings in config.toml
• OpenManus-RL: a companion project on reinforcement-learning-based tuning (such as GRPO) for LLM agents, developed with researchers from UIUC