OpenManus: The Free Open Source Manus AI Agent You Can Run Locally

Added: 2026-05-12

[00:00:00]Manis is one of those AI agents that came out of a Chinese startup called Monica. And the demo videos racked up over 1 million views in less than a day.

[00:00:10]The pitch is pretty simple. It's not just a chatbot that gives you text answers. It actually goes off and does the work for you. Once you log in, the dashboard gives you a real sense of how broad this thing is. You can ask it to build a website, develop desktop apps, create slides, run a wide research task across the internet, generate a full spreadsheet from raw data, produce visualizations. So instead of opening five different tools for five different jobs, you give Manis one prompt and it picks the right path. On the G AIA benchmark, which is the standard test for general AI agents built by Meta, Hugging Face, and the Auto GPT team, Manis reportedly scored around 86% on the basic level and over 70% on the intermediate level. That puts it ahead of OpenAI's deep research on those tiers. So, this isn't just hype. The agent actually performs. Now, the platform is hosted in China. Your prompts and outputs go through their servers and the free plan has tight credit limits. If you want to run anything heavy, you're paying. And for developers, the bigger issue is that you can't see what's happening under the hood. You can't modify the agent loop.

[00:01:24]You can't swap models, and you can't run it offline. That's the gap Open Manis is trying to fill. A small team from Metag GPT, which is another well-known agent project, built it as a fully open-source version of the same idea. They put the prototype together in about 3 hours, and it now sits on GitHub with around 56,000 stars and almost 10,000 forks. The repo is licensed under MIT, which means you can do basically anything you want with it. The way Open Manis actually works is closer to a framework than a finished product. You give it a goal in plain English and it breaks that goal into smaller steps. Then it picks the right tool for each step. Sometimes that's a browser, sometimes it's a code interpreter, sometimes it's a file system. It runs the step, looks at the result, and decides what to do next. The same loop you'd see in Manis, just exposed in code you can read and modify.

[00:02:23]One thing worth knowing is that the team also released a sister project called open manis ever s. That one is built with researchers from UIU and the goal is to use reinforcement learning specifically methods like GRPO to make the agent better at reasoning and tool use over time. So you have the main framework you can run today and a research track that's trying to push the agent quality further. Most people will just want the main one. The big honest trade-off with Open Manis is that there's no graphical interface. You run it from the terminal. You watch the steps print out. You don't get the polished dashboard that the Manis product shows. For some people, that's a deal breakaker. For others, especially developers, it's actually better because you can see exactly what the agent is doing and swap out parts you don't like.

[00:03:15]All right, let's get into installation.

[00:03:17]The official repo gives you two methods.

[00:03:20]Method one uses and method two uses UV.

[00:03:24]The team recommends method two because UV is way faster at installing Python packages and that's what we're going to use. First thing you need is UV itself.

[00:03:33]If you're on Mac or Linux, you can install it with a single curl command.

[00:03:38]After it's installed, close your terminal and reopen it so the path picks up correctly. Next, clone the repo. You run git clone command. After it pulls down, change into the open manis folder with cd open manis. Now you make the virtual environment so you don't mess up anything else on your system. To activate it on Mac or Linux, you run command. On Windows, you run.vnvm/scripts back/activate.

[00:04:06]Once your prompt shows the environment is active, you install the dependencies.

[00:04:11]The command is uvp pip install-r requirements.txt.

[00:04:16]This is where UV really pays off. What used to take a few minutes with regular pip finishes in seconds. There's also one optional step. If you want the agent to control a browser, which honestly is one of the more interesting things it can do, you run playrite install. That pulls down the browser binaries it needs. Now for the configuration part, which trips up a lot of people. Open Manis needs to know which large language model to use. Inside the config folder, there's a file called config. You copy that file and rename the copy to config.2L.

[00:04:52]You can do that in the terminal or just in your file explorer. Open config.2L in VS Code or whatever editor you like.

[00:05:01]You'll see a section for the main language model and an optional section for vision. The defaults point to GPT40 through the OpenAI API. If you have an OpenAI key, just paste it in. If you don't, you have options. You can switch the base URL and model name to use Anthropics Claude or Google's Gemini or DeepSeek or even local models through Alama. The config takes any provider that follows the Open AI API format, which at this point is most of them. For people who don't want to pay anything, services like Grock and Hyperbolic offer free tiers with rate limits and they work fine for testing. Just be aware that the smaller free models will give you worse results than something like GPT40 or Claude. The agent makes a lot of decisions, so the quality of the underlying model really shows up in the final output. After you save the config, you're ready to run it. The simplest entry point is Python main. py that gives you the standard open manis agent.

[00:06:05]There's also python run_mcp.py if you want to use the mcp tool version which lets the agent talk to external tools through the model context protocol. And there's python run_flow.py for the multi-agent setup which they call unstable but works for things like data analysis where you want a planner agent and a worker agent talking to each other. When you run main.p py it asks you to type in your idea. So I tried something practical. I asked it to build a small habit tracker as an HTML page with a clean design and the ability to mark days as complete. It went through the planning phase, decided it needed to write a single HTML file with embedded CSS and JavaScript, generated the code, saved it to the workspace folder, and then opened it to verify it worked. The whole thing took maybe four or 5 minutes. The output wasn't perfect. The self-correction loop is the part that actually makes these agent frameworks interesting. You're not just getting oneshot output. You're getting something that checks its own work. Compared to the actual Manis product, Open Manis is slower and less polished. The original has a smooth interface, runs on a hosted backend, and the underlying agent has clearly been tuned a lot. Open Manis feels more like a kit. You assemble it, you configure it, you watch it run in a terminal, but it's free. It works on your own API keys, and you own all the data. For people who care about privacy, that last part matters. With Manis, your prompts and outputs go through their servers. With Open Manis running locally, the only thing leaving your machine is the API call to whatever language model you picked. There are a few practical limitations to know about.

[00:07:55]The agent can sometimes get stuck in loops, especially with cheaper models.

[00:07:59]It can also burn through API credits faster than you'd expect because each step is a full LLM call. I'd recommend setting a spending cap on whatever provider you use just to be safe, and the team is clear that this is still early. New features and fixes land almost every week. If you want to follow along with the project, the GitHub issues page is pretty active. And there's a Discord linked in the readme where the maintainers actually respond.

[00:08:27]They also have a paper now and a hugging face demo space if you want to try a hosted version before installing anything. That's the full picture on Open Manis. It's not a polished product.

[00:08:39]It's a working open-source agent framework that gives you most of what Manis does on your own machine with your own keys and no monthly credit cap. For developers and people who want to learn how these agents actually function under the hood, it's worth setting up. For everyone else, it's a good signal of where this whole agent category is heading. Closed products generate the hype, but the open-source clones are the ones that end up getting actually used.

[00:09:07]All right, so that's it from the video and I hope you enjoyed it.