🤖 OpenManus — Local-First Autonomous AI Action Engine

OpenManus is an open-source, local-first autonomous AI action engine designed to orchestrate complex reasoning, coding, and web-browsing workflows. Inspired by the capabilities of next-generation AI assistants like Manus, OpenManus allows developers and power users to deploy agentic loops directly on their machines. The engine handles long-running goals by writing code, executing scripts in an isolated sandbox, and navigating the web, while presenting a real-time interactive remote browser screen. This stream enables users to observe the agent's progress and actively guide the session by interacting with the browser interface in real time.

---

📽️ Demo & Screenshots

(Insert GIF demonstrating live browser control, terminal execution, and settings toggling here)

[DEMO GIF PLACEHOLDER]

---

🌟 Capabilities & Features

Select any section below to learn how it works under the hood:

🎭 3 Specialized Agent Roles

OpenManus optimizes performance by offering three specialized agent personas that can be selected directly in the chat panel depending on the task:

• OpenManus (Autonomous Orchestrator): The default general-purpose planner that handles standard tasks, tool selection, and reasoning.
• CoderAgent (Software Developer): Focuses on code generation, script execution, environment resolution, and math/data tasks inside the Docker sandbox.
• BrowserAgent (Web Researcher): Optimizes search, web browsing, scraping, and remote interaction.

🔒 Persistent Docker Sandbox & Docker-in-Docker

All code execution (Python, Node.js, Shell scripts) happens inside a persistent, secure Docker sandbox container:

• In-Memory Tar Streams: Pushes scripts and assets into the sandbox via tar stream archives to avoid local file volume mount issues.
• Persistent Runtimes: The container remains running during the session to preserve installed packages (e.g. if the agent runs pip install pandas in step 1, it remains available in subsequent steps).
• Docker-in-Docker: Mounts the host's /var/run/docker.sock to allow the agent to run and manage its own nested Docker containers for complex multi-container workflows.
• Resource Constraints: Enforces a 512MB RAM cap, 80% CPU limit, and a PID limit of 128 to prevent runaway processes.

💻 Coding, File Creation & Live Hosted URLs

The agent has full read/write access to the workspace directory to build complete projects:

• Workspace File Management: Reads, writes, modifies, and deletes files directly within the local workspace.
• Live Local Servers: The agent can spin up development servers (like React Vite, Express, or Python Flask) inside the environment.
• Link Recognition: If a local server is started, it is automatically exposed on localhost and rendered as a live, clickable host URL in the chat and execution panel.

🛠️ Workflows & Skills Store

OpenManus features a database-backed Skills Store that allows the agent to save and reuse complex workflows:

• Workflow Distillation: The agent can bundle a multi-step workflow (such as scraping a site, extracting data, and formatting a report) into a reusable skill.
• Interactive Execution: Call upon saved skills inside chat sessions to perform automated routines.

🧠 Long-Term Postgres Memory

• Persistent Facts: Remembers folders, habits, preferences, or credentials across separate chat sessions.
• LLM Consolidation: Periodically merges duplicate memories into a compact list using LLM consolidation to minimize context-window footprint.

⚡ Context Compacting & AI Summary

• Configurable Summarization Threshold (in characters).
• Automatically compacts old messages and logs when chat size exceeds the limit, maintaining the agent's core context window.

📊 Agent Workspace & Session Diagnostics

The Diagnostics & Inspector panel on the right side of the interface provides deep visibility into the agent's active session:

• Timeline: Visualization of task steps and agent execution timeline.
• Thoughts: The agent's raw chain-of-thought (thinking blocks) parsed from <thinking> tags.
• Logs: Raw logs showing command outputs and terminal execution history.
• JSON: Full session JSON state inspector.
• Browser: Live interactive browser player (CDP stream canvas player).
• Files: Workspace file explorer allowing you to read/inspect files created by the agent in real time.
• Prompt: Inspection of the dynamic System Prompt and injected messages history.

---

🛠️ Prerequisites

Make sure you have these installed:

• Node.js (version 20.x or higher)
• pnpm package manager
• PostgreSQL running locally (or via Docker Compose)
• Docker Desktop (with "Expose daemon on tcp://localhost:2375" checked in settings on Windows)
• Ollama (Optional): If you want to use local models (like qwen2.5:7b or gemma2). OpenManus also natively supports Groq and OpenAI cloud providers out-of-the-box.

---

🚀 Quick Start & Installation

1. Database Setup (Docker Compose)

Run the pre-configured Postgres database container:

# Starts PostgreSQL (exposed on port 5432) and pgAdmin (on port 5050)
pnpm run docker:up

Note: The schema is automatically applied and seeded with initial skills on first boot.

2. Environment Variables

Copy the template and adjust values:

cp .env.example .env

Inside your .env, specify your database credentials:

POSTGRES_HOST=localhost
POSTGRES_PORT=5432
POSTGRES_DB=openmanus
POSTGRES_USER=postgres
POSTGRES_PASSWORD=postgres

3. Install Dependencies

OpenManus is structured as a pnpm monorepo workspace. Running install at the root handles both backend and frontend automatically:

# Option A: Automatic installation of all dependencies (root & frontend)
pnpm install

Alternatively, you can install them separately if needed:

# Option B: Separate installations
pnpm install                     # Install backend/root dependencies
pnpm --filter frontend install   # Install frontend dependencies

4. Install Browser Binary

Installs the headless Chrome binary needed for web browsing:

pnpm run browser:install

5. Launch the Platform

Start both the Express backend and Vite frontend concurrently:

pnpm run dev:all

• Frontend Client: Visit http://localhost:5173.
• Backend API: Running at http://localhost:3000.

---

📄 License

This project is open-source software licensed under the MIT License.