Cua is an open-source framework that enables AI agents to control full operating systems within high-performance, lightweight virtual containers. It delivers up to 97% native CPU speed on Apple Silicon and works with any language models.
When we started building Computer-Use AI agents we quickly realised that for them to be genuinely useful, they must interact with your system's native applications. But giving full access to your host device is risky - and not ideal. How can AI agents interact with operating systems, desktop applications, and browsers without jeopardizing security or sacrificing performance?
Today, we're thrilled to introduce Cua: an open-source framework that lets AI agents control full operating systems within secure, high-performance virtual containers.
Cua brings together three powerful components into one unified framework stack:
Lume - Our high-performance Virtualization layer leverages Apple's Virtualization.Framework to spin up macOS or Linux environments.
Computer-Use Interface - This layer lets any process or AI agent - no matter which framework it's built on (like AutoGen, LangGraph, etc.) - see the screen and interact with Lume VMs just like a human would. It supports actions such as clicking, typing, scrolling, and extracting the visual accessibility tree.
Computer-Use Agent - A general agent framework to run intelligent workflows within these VMs. It's compatible with popular LLM providers like OpenAI, Anthropic, UI-Tars, as well as open-source tools like Ollama, LM Studio, and OpenRouter.
AI agents built with Cua operates within a virtual environment to:
Navigate and interact with any application's interface
Read screen content and perform keyboard/mouse actions
Switch between applications and self-debug when needed
Operate in a secure sandbox with controlled file access
All of this happens inside a fully isolated virtual container, ensuring your host system, files, and sensitive data remain completely safe - while you continue using your device uninterrupted.
But we didn’t stop there. We’ve also launched a cloud platform for Cua, enabling developers to spin up and control these remote containers via API - no need to run anything locally. This makes it easy to scale and deploy secure, cloud-hosted agents that interact with real OS environments - perfect for testing, remote execution, or powering AI-first products.
From CAD software to Photoshop, from Tableau to legacy tools with no APIs, Cua opens up entirely new possibilities for human-AI collaboration.
We built Cua because we believe AI should enhance the tools we already use - not replace them. It should operate alongside us, securely, transparently, and on our terms.
The journey is just beginning, and we can’t wait to see what you’ll build!
@antonyo_demydov Hey Antonyo - just to clarify, when you say "custom GPT," do you mean setting up a custom prompt connected to different models? If so, you can absolutely do that! Let me know if you meant something else.
@pritraveler Hey Yong, thanks for the awesome comment! 🙌
You nailed it - while other tools are still in the "few clicks" era, Cua is pushing toward true zero-click automation. The whole idea is to create these super secure, isolated environments where AI agents can just do their thing without constant human babysitting :)
This could redefine cross-platform automation! Love the Apple Virtualization.Framework integration for macOS devs! How does Cua handle permissions? Will it need human's intervention for kernel-level actions?
@desmond_ren1 Thanks Desmond! For regular dev stuff (coding, testing, running apps), Cua handles permissions automatically within the VM - no need to worry! But for hardcore kernel-level actions or loading custom kexts, you'll need to step in manually. Cua intentionally doesn't automate these high-risk operations to protect your systems!
Cua feels like a foundational step for the next generation of AI agents—giving them full OS control in lightweight containers is a powerful move for scalable, real-world automation.
@nicoleastor That's great to hear! Really excited that you're interested in this secure approach for AI app interaction. I'm curious - what specific use case are you looking to implement? What kind of apps are you hoping to let AI use?
Happy to clarify the difference: Cua is the complete framework that enables AI agents to control operating systems in secure containers. Lume is actually one component of the Cua stack - specifically the virtualization layer that leverages Apple's Virtualization.Framework to run the high-performance VMs. Our users typically use both together - Lume provides the virtualization foundation, while the full Cua framework gives you the agent capabilities.
Cua
When we started building Computer-Use AI agents we quickly realised that for them to be genuinely useful, they must interact with your system's native applications. But giving full access to your host device is risky - and not ideal. How can AI agents interact with operating systems, desktop applications, and browsers without jeopardizing security or sacrificing performance?
Today, we're thrilled to introduce Cua: an open-source framework that lets AI agents control full operating systems within secure, high-performance virtual containers.
Cua brings together three powerful components into one unified framework stack:
Lume - Our high-performance Virtualization layer leverages Apple's Virtualization.Framework to spin up macOS or Linux environments.
Computer-Use Interface - This layer lets any process or AI agent - no matter which framework it's built on (like AutoGen, LangGraph, etc.) - see the screen and interact with Lume VMs just like a human would. It supports actions such as clicking, typing, scrolling, and extracting the visual accessibility tree.
Computer-Use Agent - A general agent framework to run intelligent workflows within these VMs. It's compatible with popular LLM providers like OpenAI, Anthropic, UI-Tars, as well as open-source tools like Ollama, LM Studio, and OpenRouter.
AI agents built with Cua operates within a virtual environment to:
Navigate and interact with any application's interface
Read screen content and perform keyboard/mouse actions
Switch between applications and self-debug when needed
Operate in a secure sandbox with controlled file access
All of this happens inside a fully isolated virtual container, ensuring your host system, files, and sensitive data remain completely safe - while you continue using your device uninterrupted.
But we didn’t stop there. We’ve also launched a cloud platform for Cua, enabling developers to spin up and control these remote containers via API - no need to run anything locally. This makes it easy to scale and deploy secure, cloud-hosted agents that interact with real OS environments - perfect for testing, remote execution, or powering AI-first products.
From CAD software to Photoshop, from Tableau to legacy tools with no APIs, Cua opens up entirely new possibilities for human-AI collaboration.
We built Cua because we believe AI should enhance the tools we already use - not replace them. It should operate alongside us, securely, transparently, and on our terms.
The journey is just beginning, and we can’t wait to see what you’ll build!
Check out our website at cua.sh
GitHub repo at github.com/trycua/cua
@pupposandro can I run my own custom GPT agent inside Cua?
Cua
@antonyo_demydov Hey Antonyo - just to clarify, when you say "custom GPT," do you mean setting up a custom prompt connected to different models? If so, you can absolutely do that! Let me know if you meant something else.
Den
@pupposandro @antonyo_demydov I was able to pretty easily
Prit
While others say 'Finish with a few clicks', Cua will make an environment that users even don't need a single click!
Congratulations :)
Cua
@pritraveler Hey Yong, thanks for the awesome comment! 🙌
You nailed it - while other tools are still in the "few clicks" era, Cua is pushing toward true zero-click automation. The whole idea is to create these super secure, isolated environments where AI agents can just do their thing without constant human babysitting :)
Den
@pritraveler max 1 click haha!
Manna
This could redefine cross-platform automation! Love the Apple Virtualization.Framework integration for macOS devs! How does Cua handle permissions? Will it need human's intervention for kernel-level actions?
Cua
@desmond_ren1 Thanks Desmond! For regular dev stuff (coding, testing, running apps), Cua handles permissions automatically within the VM - no need to worry! But for hardcore kernel-level actions or loading custom kexts, you'll need to step in manually. Cua intentionally doesn't automate these high-risk operations to protect your systems!
Den
@desmond_ren1 @pupposandro Wow you have VM-level permissions? Crazy
Cua feels like a foundational step for the next generation of AI agents—giving them full OS control in lightweight containers is a powerful move for scalable, real-world automation.
Cua
@supa_l thanks for the comment - that's exactly what we envisioned while building Cua!
Den
@supa_l it's going to be a massive unlock imo.
YouMind
Finally, a secure way to let AI use my apps! Been waiting for something like this forever!
Cua
@nicoleastor That's great to hear! Really excited that you're interested in this secure approach for AI app interaction. I'm curious - what specific use case are you looking to implement? What kind of apps are you hoping to let AI use?
Den
@nicoleastor Can confirm it works better than expected, lol
How does running full virtual machines for AI agents affect performance and scalability on computers with less memory and processing power?
great launch! whats the difference netween Cua and Lume? may I need both?
Cua
@mark_hokil Thanks for the kind words Mark!
Happy to clarify the difference: Cua is the complete framework that enables AI agents to control operating systems in secure containers. Lume is actually one component of the Cua stack - specifically the virtualization layer that leverages Apple's Virtualization.Framework to run the high-performance VMs. Our users typically use both together - Lume provides the virtualization foundation, while the full Cua framework gives you the agent capabilities.