Security & Trust.

ZeroSphere is BYOK (Bring Your Own Keys). When you run the agent, your code goes directly to your model provider, whether Anthropic, OpenAI, or Google, using your API key. It never touches our servers.

ZeroSphere runs on your machine. The editor, the agent, the virtual display, and the build-run-fix loop are all entirely local.

We can't leak what we don't store.

Most AI tools that interact with UIs read the DOM or the accessibility tree. That's the text layer underneath what you see, and it's exactly where prompt injection attacks live.

A malicious website, a compromised dependency, or a crafted UI element can write hidden text into the DOM that an AI reads and obeys.

Pure raw pixels. A screenshot taken by an internal capture tool, passed directly to the model as an image. No DOM. No accessibility API. No text layer. No hidden instructions.

This isn't a security feature we added. It's a consequence of how vision-based perception works, and it makes an entire class of AI agent attacks structurally impossible against ZeroSphere.

ZeroSphere's agent operates inside an isolated virtual display, which is a sandboxed screen separate from your desktop. It interacts with your app inside that boundary. You control which applications it can launch and interact with. Every action is visible in real time. Nothing runs hidden.

The agent asks before it acts outside expected scope. You can grant permissions per action type, revoke them mid-session, and see a full log of everything it did. No black boxes.