Login
🐫 CAMEL is an open-source community dedicated to finding the scaling laws of agents. We believe that studying these agents on a large scale offers valuable insights into their behaviors, capabilities, and potential risks. To facilitate research in this field, we implement and support various types of agents, tasks, prompts, models, and simulated environments.
The framework is designed to support systems with millions of agents, ensuring efficient coordination, communication, and resource management at scale.
Agents maintain stateful memory, enabling them to perform multi-step interactions with environments and efficiently tackle sophisticated tasks.
Every line of code and comment serves as a prompt for agents. Code should be written clearly and readably, ensuring both humans and agents can interpret it effectively.
We are a community-driven research collective comprising over 100 researchers dedicated to advancing frontier research in Multi-Agent Systems. Researchers worldwide choose CAMEL for their studies based on the following reasons.
| ✅ | Large-Scale Agent System | Simulate up to 1M agents to study emergent behaviors and scaling laws in complex, multi-agent environments. |
| ✅ | Dynamic Communication | Enable real-time interactions among agents, fostering seamless collaboration for tackling intricate tasks. |
| ✅ | Stateful Memory | Equip agents with the ability to retain and leverage historical context, improving decision-making over extended interactions. |
| ✅ | Support for Multiple Benchmarks | Utilize standardized benchmarks to rigorously evaluate agent performance, ensuring reproducibility and reliable comparisons. |
| ✅ | Support for Different Agent Types | Work with a variety of agent roles, tasks, models, and environments, supporting interdisciplinary experiments and diverse research applications. |
| ✅ | Data Generation and Tool Integration | Automate the creation of large-scale, structured datasets while seamlessly integrating with multiple tools, streamlining synthetic data generation and research workflows. |
Installing CAMEL is a breeze thanks to its availability on PyPI. Simply open your terminal and run:
FreshRSS pip install camel-ai
This example demonstrates how to create a ChatAgent using the CAMEL framework and perform a search query using DuckDuckGo.
bash pip install 'camel-ai[web_tools]'
bash export OPENAI_API_KEY='your_openai_api_key'
```python from camel.models import ModelFactory from camel.types import ModelPlatformType, ModelType from camel.agents import ChatAgent from camel.toolkits import SearchToolkit
model = ModelFactory.create( model_platform=ModelPlatformType.OPENAI, model_type=ModelType.GPT_4O, model_config_dict={"temperature": 0.0}, )
search_tool = SearchToolkit().search_duckduckgo
agent = ChatAgent(model=model, tools=[search_tool])
response_1 = agent.step("What is CAMEL-AI?") print(response_1.msgs[0].content) # CAMEL-AI is the first LLM (Large Language Model) multi-agent framework # and an open-source community focused on finding the scaling laws of agents. # ...
response_2 = agent.step("What is the Github link to CAMEL framework?") print(response_2.msgs[0].content) # The GitHub link to the CAMEL framework is # https://github.com/camel-ai/camel. ```
For more detailed instructions and additional configuration options, check out the installation section.
After running, you can explore our CAMEL Tech Stack and Cookbooks at docs.camel-ai.org to build powerful multi-agent systems.
We provide a demo showcasing a conversation between two ChatGPT agents playing roles as a python programmer and a stock trader collaborating on developing a trading bot for stock market.
Explore different types of agents, their roles, and their applications.
Please reach out to us on CAMEL discord if you encounter any issue set up CAMEL.
Core components and utilities to build, operate, and enhance CAMEL-AI agents and societies.
| Module | Description |
|---|---|
| Agents | Core agent architectures and behaviors for autonomous operation. |
| Agent Societies | Components for building and managing multi-agent systems and collaboration. |
| Data Generation | Tools and methods for synthetic data creation and augmentation. |
| Models | Model architectures and customization options for agent intelligence. |
| Tools | Tools integration for specialized agent tasks. |
| Memory | Memory storage and retrieval mechanisms for agent state management. |
| Storage | Persistent storage solutions for agent data and states. |
| Benchmarks | Performance evaluation and testing frameworks. |
| Interpreters | Code and command interpretation capabilities. |
| Data Loaders | Data ingestion and preprocessing tools. |
| Retrievers | Knowledge retrieval and RAG components. |
| Runtime | Execution environment and process management. |
| Human-in-the-Loop | Interactive components for human oversight and intervention. |
| --- |
We believe that studying these agents on a large scale offers valuable insights into their behaviors, capabilities, and potential risks.
Explore our research projects:
Research with US
We warmly invite you to use CAMEL for your impactful research.
Rigorous research takes time and resources. We are a community-driven research collective with 100+ researchers exploring the frontier research of Multi-agent Systems. Join our ongoing projects or test new ideas with us, reach out via email for more information.
For more details, please see our Models Documentation.
Data (Hosted on Hugging Face)
| Dataset | Chat format | Instruction format | Chat format (translated) |
|---|---|---|---|
| AI Society | Chat format | Instruction format | Chat format (translated) |
| Code | Chat format | Instruction format | x |
| Math | Chat format | x | x |
| Physics | Chat format | x | x |
| Chemistry | Chat format | x | x |
| Biology | Chat format | x | x |
| Dataset | Instructions | Tasks |
|---|---|---|
| AI Society | Instructions | Tasks |
| Code | Instructions | Tasks |
| Misalignment | Instructions | Tasks |
Practical guides and tutorials for implementing specific functionalities in CAMEL-AI agents and societies.
| Cookbook | Description |
|---|---|
| Creating Your First Agent | A step-by-step guide to building your first agent. |
| Creating Your First Agent Society | Learn to build a collaborative society of agents. |
| Message Cookbook | Best practices for message handling in agents. |
| Cookbook | Description |
|---|---|
| Tools Cookbook | Integrating tools for enhanced functionality. |
| Memory Cookbook | Implementing memory systems in agents. |
| RAG Cookbook | Recipes for Retrieval-Augmented Generation. |
| Graph RAG Cookbook | Leveraging knowledge graphs with RAG. |
| Track CAMEL Agents with AgentOps | Tools for tracking and managing agents in operations. |
| Cookbook | Description |
|---|---|
| Data Generation with CAMEL and Finetuning with Unsloth | Learn how to generate data with CAMEL and fine-tune models effectively with Unsloth. |
| Data Gen with Real Function Calls and Hermes Format | Explore how to generate data with real function calls and the Hermes format. |
| CoT Data Generation and Upload Data to Huggingface | Uncover how to generate CoT data with CAMEL and seamlessly upload it to Huggingface. |
| CoT Data Generation and SFT Qwen with Unsolth | Discover how to generate CoT data using CAMEL and SFT Qwen with Unsolth, and seamlessly upload your data and model to Huggingface. |
| Cookbook | Description |
|---|---|
| Role-Playing Scraper for Report & Knowledge Graph Generation | Create role-playing agents for data scraping and reporting. |
| Create A Hackathon Judge Committee with Workforce | Building a team of agents for collaborative judging. |
| Dynamic Knowledge Graph Role-Playing: Multi-Agent System with dynamic, temporally-aware knowledge graphs | Builds dynamic, temporally-aware knowledge graphs for financial applications using a multi-agent system. It processes financial reports, news articles, and research papers to help traders analyze data, identify relationships, and uncover market insights. The system also utilizes diverse and optional element node deduplication techniques to ensure data integrity and optimize graph structure for financial decision-making. |
| Customer Service Discord Bot with Agentic RAG | Learn how to build a robust customer service bot for Discord using Agentic RAG. |
| Customer Service Discord Bot with Local Model | Learn how to build a robust customer service bot for Discord using Agentic RAG which supports local deployment. |
| Cookbook | Description |
|---|---|
| Video Analysis | Techniques for agents in video data analysis. |
| 3 Ways to Ingest Data from Websites with Firecrawl | Explore three methods for extracting and processing data from websites using Firecrawl. |
| Create AI Agents that work with your PDFs | Learn how to create AI agents that work with your PDFs using Chunkr and Mistral AI. |
For those who'd like to contribute code, we appreciate your interest in contributing to our open-source initiative. Please take a moment to review our contributing guidelines to get started on a smooth collaboration journey.🚀
We also welcome you to help CAMEL grow by sharing it on social media, at events, or during conferences. Your support makes a big difference!
For more information please contact camel-ai@eigent.ai
GitHub Issues: Report bugs, request features, and track development. Submit an issue
Discord: Get real-time support, chat with the community, and stay updated. Join us
X (Twitter): Follow for updates, AI insights, and key announcements. Follow us
Ambassador Project: Advocate for CAMEL-AI, host events, and contribute content. Learn more
@inproceedings{li2023camel,
title={CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society},
author={Li, Guohao and Hammoud, Hasan Abed Al Kader and Itani, Hani and Khizbullin, Dmitrii and Ghanem, Bernard},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
year={2023}
}
Special thanks to Nomic AI for giving us extended access to their data set exploration tool (Atlas).
We would also like to thank Haya Hammoud for designing the initial logo of our project.
We implemented amazing research ideas from other works for you to build, compare and customize your agents. If you use any of these modules, please kindly cite the original works: - TaskCreationAgent, TaskPrioritizationAgent and BabyAGI from Nakajima et al.: Task-Driven Autonomous Agent. [Example]
PersonaHub from Tao Ge et al.: Scaling Synthetic Data Creation with 1,000,000,000 Personas. [Example]
Self-Instruct from Yizhong Wang et al.: SELF-INSTRUCT: Aligning Language Models with Self-Generated Instructions. [Example]
The source code is licensed under Apache 2.0.
WormGPT, a private new chatbot service advertised as a way to use Artificial Intelligence (AI) to write malicious software without all the pesky prohibitions on such activity enforced by the likes of ChatGPT and Google Bard, has started adding restrictions of its own on how the service can be used. Faced with customers trying to use WormGPT to create ransomware and phishing scams, the 23-year-old Portuguese programmer who created the project now says his service is slowly morphing into “a more controlled environment.”
Image: SlashNext.com.
The large language models (LLMs) made by ChatGPT parent OpenAI or Google or Microsoft all have various safety measures designed to prevent people from abusing them for nefarious purposes — such as creating malware or hate speech. In contrast, WormGPT has promoted itself as a new, uncensored LLM that was created specifically for cybercrime activities.
WormGPT was initially sold exclusively on HackForums, a sprawling, English-language community that has long featured a bustling marketplace for cybercrime tools and services. WormGPT licenses are sold for prices ranging from 500 to 5,000 Euro.
“Introducing my newest creation, ‘WormGPT,’ wrote “Last,” the handle chosen by the HackForums user who is selling the service. “This project aims to provide an alternative to ChatGPT, one that lets you do all sorts of illegal stuff and easily sell it online in the future. Everything blackhat related that you can think of can be done with WormGPT, allowing anyone access to malicious activity without ever leaving the comfort of their home.”
WormGPT’s core developer and frontman “Last” promoting the service on HackForums. Image: SlashNext.
In July, an AI-based security firm called SlashNext analyzed WormGPT and asked it to create a “business email compromise” (BEC) phishing lure that could be used to trick employees into paying a fake invoice.
“The results were unsettling,” SlashNext’s Daniel Kelley wrote. “WormGPT produced an email that was not only remarkably persuasive but also strategically cunning, showcasing its potential for sophisticated phishing and BEC attacks.”
SlashNext asked WormGPT to compose this BEC phishing email. Image: SlashNext.
A review of Last’s posts on HackForums over the years shows this individual has extensive experience creating and using malicious software. In August 2022, Last posted a sales thread for “Arctic Stealer,” a data stealing trojan and keystroke logger that he sold there for many months.
“I’m very experienced with malwares,” Last wrote in a message to another HackForums user last year.
Last has also sold a modified version of the information stealer DCRat, as well as an obfuscation service marketed to malicious coders who sell their creations and wish to insulate them from being modified or copied by customers.
Shortly after joining the forum in early 2021, Last told several different Hackforums users his name was Rafael and that he was from Portugal. HackForums has a feature that allows anyone willing to take the time to dig through a user’s postings to learn when and if that user was previously tied to another account.
That account tracing feature reveals that while Last has used many pseudonyms over the years, he originally used the nickname “ruiunashackers.” The first search result in Google for that unique nickname brings up a TikTok account with the same moniker, and that TikTok account says it is associated with an Instagram account for a Rafael Morais from Porto, a coastal city in northwest Portugal.
Reached via Instagram and Telegram, Morais said he was happy to chat about WormGPT.
“You can ask me anything,” Morais said. “I’m an open book.”
Morais said he recently graduated from a polytechnic institute in Portugal, where he earned a degree in information technology. He said only about 30 to 35 percent of the work on WormGPT was his, and that other coders are contributing to the project. So far, he says, roughly 200 customers have paid to use the service.
“I don’t do this for money,” Morais explained. “It was basically a project I thought [was] interesting at the beginning and now I’m maintaining it just to help [the] community. We have updated a lot since the release, our model is now 5 or 6 times better in terms of learning and answer accuracy.”
WormGPT isn’t the only rogue ChatGPT clone advertised as friendly to malware writers and cybercriminals. According to SlashNext, one unsettling trend on the cybercrime forums is evident in discussion threads offering “jailbreaks” for interfaces like ChatGPT.
“These ‘jailbreaks’ are specialised prompts that are becoming increasingly common,” Kelley wrote. “They refer to carefully crafted inputs designed to manipulate interfaces like ChatGPT into generating output that might involve disclosing sensitive information, producing inappropriate content, or even executing harmful code. The proliferation of such practices underscores the rising challenges in maintaining AI security in the face of determined cybercriminals.”
Morais said they have been using the GPT-J 6B model since the service was launched, although he declined to discuss the source of the LLMs that power WormGPT. But he said the data set that informs WormGPT is enormous.
“Anyone that tests wormgpt can see that it has no difference from any other uncensored AI or even chatgpt with jailbreaks,” Morais explained. “The game changer is that our dataset [library] is big.”
Morais said he began working on computers at age 13, and soon started exploring security vulnerabilities and the possibility of making a living by finding and reporting them to software vendors.
“My story began in 2013 with some greyhat activies, never anything blackhat tho, mostly bugbounty,” he said. “In 2015, my love for coding started, learning c# and more .net programming languages. In 2017 I’ve started using many hacking forums because I have had some problems home (in terms of money) so I had to help my parents with money… started selling a few products (not blackhat yet) and in 2019 I started turning blackhat. Until a few months ago I was still selling blackhat products but now with wormgpt I see a bright future and have decided to start my transition into whitehat again.”
WormGPT sells licenses via a dedicated channel on Telegram, and the channel recently lamented that media coverage of WormGPT so far has painted the service in an unfairly negative light.
“We are uncensored, not blackhat!” the WormGPT channel announced at the end of July. “From the beginning, the media has portrayed us as a malicious LLM (Language Model), when all we did was use the name ‘blackhatgpt’ for our Telegram channel as a meme. We encourage researchers to test our tool and provide feedback to determine if it is as bad as the media is portraying it to the world.”
It turns out, when you advertise an online service for doing bad things, people tend to show up with the intention of doing bad things with it. WormGPT’s front man Last seems to have acknowledged this at the service’s initial launch, which included the disclaimer, “We are not responsible if you use this tool for doing bad stuff.”
But lately, Morais said, WormGPT has been forced to add certain guardrails of its own.
“We have prohibited some subjects on WormGPT itself,” Morais said. “Anything related to murders, drug traffic, kidnapping, child porn, ransomwares, financial crime. We are working on blocking BEC too, at the moment it is still possible but most of the times it will be incomplete because we already added some limitations. Our plan is to have WormGPT marked as an uncensored AI, not blackhat. In the last weeks we have been blocking some subjects from being discussed on WormGPT.”
Still, Last has continued to state on HackForums — and more recently on the far more serious cybercrime forum Exploit — that WormGPT will quite happily create malware capable of infecting a computer and going “fully undetectable” (FUD) by virtually all of the major antivirus makers (AVs).
“You can easily buy WormGPT and ask it for a Rust malware script and it will 99% sure be FUD against most AVs,” Last told a forum denizen in late July.
Asked to list some of the legitimate or what he called “white hat” uses for WormGPT, Morais said his service offers reliable code, unlimited characters, and accurate, quick answers.
“We used WormGPT to fix some issues on our website related to possible sql problems and exploits,” he explained. “You can use WormGPT to create firewalls, manage iptables, analyze network, code blockers, math, anything.”
Morais said he wants WormGPT to become a positive influence on the security community, not a destructive one, and that he’s actively trying to steer the project in that direction. The original HackForums thread pimping WormGPT as a malware writer’s best friend has since been deleted, and the service is now advertised as “WormGPT – Best GPT Alternative Without Limits — Privacy Focused.”
“We have a few researchers using our wormgpt for whitehat stuff, that’s our main focus now, turning wormgpt into a good thing to [the] community,” he said.
It’s unclear yet whether Last’s customers share that view.
A GPT-empowered penetration testing tool.
resources where we use it to solve HackTheBox challenge TEMPLATED (web challenge).Before installation, we recommend you to take a look at this installation video if you want to use cookie setup.
requirements.txt with pip install -r requirements.txt
config. You may follow a sample by cp config/chatgpt_config_sample.py config/chatgpt_config.py. Inspect - Network, find the connections to the ChatGPT session page.https://chat.openai.com/api/auth/session and paste it into the cookie field of config/chatgpt_config.py. (You may use Inspect->Network, find session and copy the cookie field in request_headers to https://chat.openai.com/api/auth/session)userAgent with your user agent.chatgpt_config.py.python3 test_connection.py. You should see some sample conversation with ChatGPT. 1. You're connected with ChatGPT Plus cookie.
To start PentestGPT, please use <python3 main.py --reasoning_model=gpt-4>
## Test connection for OpenAI api (GPT-4)
2. You're connected with OpenAI API. You have GPT-4 access. To start PentestGPT, please use <python3 main.py --reasoning_model=gpt-4 --useAPI>
## Test connection for OpenAI api (GPT-3.5)
3. You're connected with OpenAI API. You have GPT-3.5 access. To start PentestGPT, please use <python3 main.py --reasoning_model=gpt-3.5-turbo --useAPI>
https://chat.openai.com/backend-api/conversations. Please submit an issue if you encounter any problem.python3 main.py --args. --reasoning_model is the reasoning model you want to use.--useAPI is whether you want to use OpenAI API.test_connection.py, which are: python3 main.py --reasoning_model=gpt-4python3 main.py --reasoning_model=gpt-4 --useAPIpython3 main.py --reasoning_model=gpt-3.5-turbo --useAPIhelp: show the help message.next: key in the test execution result and get the next step.more: let PentestGPT to explain more details of the current step. Also, a new sub-task solver will be created to guide the tester.todo: show the todo list.discuss: discuss with the PentestGPT.google: search on Google. This function is still under development.quit: exit the tool and save the output as log file (see the reporting section below).TAB to autocomplete the commands.ENTER to select the item. Similarly, use <SHIFT + right arrow> to confirm selection.more, users can execute more commands to investigate into a specific problem: help: show the help message.brainstorm: let PentestGPT brainstorm on the local task for all the possible solutions.discuss: discuss with PentestGPT about this local task.google: search on Google. This function is still under development.continue: exit the subtask and continue the main testing session.logs folder (if you quit with quit command).python3 utils/report_generator.py <log file>. A sample report sample_pentestGPT_log.txt is also uploaded.Contributions are what make the open source community such an amazing place to learn, inspire, and create. Any contributions you make are greatly appreciated.
If you have a suggestion that would make this better, please fork the repo and create a pull request. You can also simply open an issue with the tag "enhancement". Don't forget to give the project a star! Thanks again!
git checkout -b feature/AmazingFeature)git commit -m 'Add some AmazingFeature')git push origin feature/AmazingFeature)Distributed under the MIT License. See LICENSE.txt for more information.
Gelei Deng - gelei.deng@ntu.edu.sg
