"Don't rob.ManusInvitation code now.Open SourceA version is on the way." This is the good news brought by CAMEL AI, a domestic open source platform, early this morning.
Their open source replica of Manus is called "OWL", and its performance on GAIA Benchmark reaches 57.7%, surpassing the performance of 55.15% of Open Deep Research proposed by Huggingface, and becoming the GAIA performance ceiling in the open source world. (GAIA is a benchmark test for evaluating the capabilities of general-purpose AI assistants, and Manus has also done related tests.)
GitHub:https://github.com/camel-ai/owl
In order for the OWL to evolve into a true fully automated multi Agent Working as a god, CAMEL AI also did a wave of wild Reverse Engineering (reverse engineering) on Manus's technology line and started a deep replica program along the way.
In this program, they broke down the core Manus workflow into 6 steps:
The project can be said to be the technical framework, workflow, core capabilities are all open, the code is all on GitHub. Interested readers can follow up by themselves.
Regarding Manus, CAMEL's initiator, Mr. Gary Lee, commented, "We have been proposing the world's first ChatGPT Multi-agent framework for almost two years now, and it's a pleasant surprise to see Monica's team turning the Multi-agent technology into a product. Monica team's Manus system design is actually very clever, especially their file system to achieve Context management and persistence, the use of the Ubuntu virtual machine command line to achieve flexible tools is very worth learning engineering ideas, we are still in the open source reproduction! We're also looking forward to the Post Training model they've promised to open source!
In addition to Web and command line control, we are very confident that computers, cell phones, robots, vehicles and other cross-platform control will become possible, then it is the real Cyber workforce, because one of the biggest advantages of OS-side Agent is that it is easier to get Verifiable reward signals, the future of Reinforcement Learning RL can be trained up to fly, perhaps in the In the future, Reinforcement Learning RL can be trained by leaps and bounds, and maybe there will be an Aha moment like DeepSeek R1 Zero in the OS/Web Agent domain within a year! It's a pity that we can't go to RL with limited resources, and we're looking forward to doing open source exploration together!
OpenManus is another open source replica of Manus besides OWL that allows users to create their own AI intelligences without an invitation code. Its usage includes setting up a new conda environment, cloning the repository, installing dependencies and configuring API keys (such as OpenAI's API) in the config.toml file. Users can enter ideas into the terminal and interact with the intelligences, which will use the LLM API to process and generate responses.
Project address:https://github.com/mannaandpoem/OpenManus
Among other things, there are some differences. As we mentioned earlier, Manus has a cloud-based computer of its own, while OpenManus lets the Agent operate your computer directly.
The MetaGPT team says the current project is still in development, with planned improvements including better planning, real-time demos, replay capabilities, enhanced learning to fine-tune the model and full benchmarking.
In conclusion, I would like to say that I expect Agent to surprise us more this year and become a real usable product.
