Imagine having an AI company that runs 24 hours a day, no HR, no managers, no overtime pay. All you need to provide is a goal, and it will automatically build a team, assign tasks, coordinate work, and even learn from mistakes.

This isn’t science fiction. The Data Science Laboratory at the University of Hong Kong (HKUDS) has just open-sourced a framework called **OpenOPC**, short for *OpenAI-Powered Company*. What it does is simple: it turns your goal into an AI-native company.


## Three Core Mechanisms

OpenOPC is built around three key concepts: **Self-Built**, **Self-Run**, and **Self-Grown**.

### 1. Self-Built: Automatic Team Formation

When you input a goal, OpenOPC first creates an organizational chart — it identifies which roles are needed for the task and the reporting hierarchy. Then a "recruiting" agent decides whether to reuse existing AI employees (who bring experience from previous projects) or hire a new worker from the talent pool.

Experienced employees bring accumulated context, while new employees are blank slates. When a task requires a completely fresh perspective, the latter actually has an advantage.

### 2. Self-Run: Automatic Execution and Collaboration

Once the team is formed, the real challenge comes: how to get multiple AI agents to collaborate efficiently?

OpenOPC solves this with a work item state machine. Every work item has a current stage, an owner, and an "executable" status. A "manager" agent is responsible for task decomposition, assignment, and result review — it can accept work, send it back for revision, or escalate it for higher-level handling.

Key details:

- Independent work items can be executed in parallel

- Dependent work items automatically wait for pre-conditions to be completed

- When blocked, the system first attempts to resolve the issue within the team

- If the issue exceeds the team’s authority, it is automatically escalated to the human user

All of this is displayed in real time on a Kanban board interface.

### 3. Self-Grown: Learning From Every Run

This is the most interesting part. After each task is completed, OpenOPC doesn’t simply assign credit or blame to the whole team. Instead, it:

- Parses user feedback into individual performance evaluations for each AI employee

- Only updates the experience of the role that actually handled the work item

- Refines the tasks of each role into "high-signal" lessons learned

- Elevates recurring lessons into shared "playbooks" that new employees inherit on onboarding

Put simply, your AI company gets smarter the more you use it.

## What Can It Do?

OpenOPC covers nine core domains:

| Domain | Specific Use Cases |

|------|----------|

| AI Technology & Research | Model training, Agent development, LLM applications |

| Software Development | Android apps, SaaS MVP, websites, mini-programs, games |

| Financial Investment | Investment memos, market maps, due diligence |

| Sales Growth | Outbound sales, channel expansion |

| Content & Media | Video production, scripts, storyboards |

| Industry Assistants | Customer service, legal, HR, retail |

| Accounting & Finance | Bookkeeping, tax, budgeting |

| Branding & E-commerce | Brand planning, product selection, operations |

| Education & Training | Curriculum design, knowledge bases, learning management |

The project also provides several demo videos:

[Video Production Demo](https://youtu.be/XqQeTt6XvPQ)

[Investment Research Demo](https://drive.google.com/drive/folders/1T1Nl6CCE-cmbGy6sKrYML7_UnP8XID88?usp=drive_link)

[Game Prototype Demo](https://youtu.be/SVc9BvE5ohY)

## How to Use It?

Installation is straightforward, we recommend using the `uv` package manager:

```bash

# Install OpenOPC

uv pip install -e .

# Initialize configuration

uv run opc init

# Launch the browser UI

uv run opc ui

```

Then open `http://localhost:8765`, and you’ll see a complete office interface:

- **Workspace**: Main working area, the Kanban board, chat, and task details are all here

- **Office**: Visualized office space, each AI employee is displayed as a pixel character, you can see what they are working on at any time

- **Org**: Company structure management, you can create new organizations, edit roles, and recruit employees


## Notable Details

1. **Fully Open Source**: Released under the MIT license, code is hosted on [GitHub](https://github.com/HKUDS/OpenOPC)

2. **Multiple Execution Modes**: Supports Task Mode (single-agent tasks) and Company Mode (multi-agent company collaboration)

3. **External Agent Integration**: Can use Codex, Claude Code, Cursor and other tools as execution engines

n4. **Platform Integration**: Supports Feishu, Telegram, Slack, Discord, DingTalk, email, WeChat and more

5. **Security Mechanism**: Tool calls have risk grading, dangerous operations such as `rm -rf` are automatically escalated to humans for approval

## A Quick Observation

OpenOPC’s ambition goes beyond being just another Agent framework. It attempts to simulate a complete organization — with hierarchy, division of labor, and a learning mechanism. When AI companies can iterate on their own and management costs approach zero, the organizational structure of traditional companies may really need to be rethought.

Of course, the project is still in its very early stages. The roadmap lists many features still waiting to be improved: role-level skills, a more complete CLI, terminal UI, a marketplace ecosystem, etc. But the direction is correct.

Project Repository: https://github.com/HKUDS/OpenOPC