Simpleyyt/ai-manus

AI Manus is a general-purpose AI Agent system that supports running various tools and operations in a sandbox environment.

What it solves

AI Manus is a general-purpose AI agent system designed to execute complex tasks by interacting with a secure, isolated environment. It eliminates the need for fragile prompt-based JSON protocols by using native tool calling, providing a reliable way for AI to run code, browse the web, and manage files without risking the host system.

How it works

The system uses a "Plan-and-execute" flow where a PlanAct Agent manages the task. When a user starts a session, the server creates a dedicated Ubuntu Docker container (sandbox) for that task. The agent then uses native tool calls to interact with this sandbox via a set of tools (Terminal, Browser, File, Web Search, and MCP integration).

To allow users to monitor the agent's progress, the system converts VNC services from the headless browser in the sandbox to WebSockets, enabling real-time viewing and manual takeover of the browser.

Who it’s for

This project is for developers and users who want a self-hosted AI agent capable of performing autonomous browser and code execution tasks within a secure, sandboxed environment.

Highlights

  • Isolated Sandboxing: Each task gets its own Docker-based Ubuntu environment.
  • Native Tool Calling: Uses structured output tools instead of JSON-in-prompt for more reliable agent loops.
  • Real-time Monitoring: Supports real-time viewing and takeover of the browser via NoVNC.
  • Extensible Skills: Supports reusable skill packages that can be loaded and synced into the sandbox.
  • Integrated Library: A sidebar library aggregates attachments and artifacts across different sessions for easy retrieval.

Related

  • Dispatch
    Securing AI Agents with Isolated Docker Sandboxes

    An exploration of agent-sandbox, a tool designed to run AI coding agents within isolated Docker containers to prevent unauthorized system access and ensure execution safety.

  • Project
    lingyuanli/MultiGen

    A general-purpose AI agent platform for private on-premise deployment featuring a Planner+ReAct architecture and isolated Docker sandboxing for secure tool execution.

  • Project
    agent-infra/sandbox

    A unified agent sandbox environment that combines a browser, shell, and IDE tools in one Docker container with a shared filesystem for AI agents.

  • Project
    docker/labs-ai-tools-for-devs

    A framework for creating agentic AI workflows where Docker containers serve as tools and Markdown files define the logic, allowing LLMs to execute complex tasks in sandboxed environments.