What it does

AI Manus is a general-purpose agent system. When you start a conversation, the server creates a separate sandbox, an Ubuntu Docker environment with a Chrome browser and tool APIs, and a plan-and-execute agent works through the task using terminal, browser, file, web search and messaging tools. The web interface lets you watch tool activity live, including the sandbox browser, and take over, stop or interrupt a task. It also supports file upload and download, a library of attachments and artifacts across sessions, reusable skill packages, external MCP tools, user login, and English and Chinese. You deploy it yourself with Docker Compose. A minimal setup needs only an LLM service, configured with an API key, base URL and model name in an env file, and session history uses MongoDB and Redis. The project asks for models with reliable native tool calling and recommends Deepseek and GPT models. It is MIT licensed.

Capabilities

Separate Docker sandbox for each task
Terminal, browser, file, web search and messaging tools
Live viewing of tool activity with the option to take over
External MCP tool integration
Reusable skill packages
Docker Compose deployment with a model that supports tool calling

When to use it

You want an agent that can use a shell and browser on infrastructure you run
You want to try a plan-and-execute agent with MCP tools using your own model API key
You want an agent with a web interface where you can watch and step in