Custom-model computer use

Choose your model. Give it the right tools.

Connect an API provider or coding agent in Settings → Agents. Choose its model when starting work. OpenAI-compatible endpoints, including configured DeepSeek and OpenRouter connections, can use tool-calling loops; OpenCode uses the task’s authenticated MCP bridge.

Browser or isolated desktop

Browser tools expose page structure, so a tool-capable text model can work with them. Background Computer runs an isolated Linux desktop. It requires QEMU, a prepared guest image, a guest SSH key and firmware where needed. Setup reports missing prerequisites.

Visual desktop work additionally requires image input. Screenshot pixels are passed to supported API and MCP consumers, rather than only returning a file path. A text-only model is not automatically capable of visual control.

You retain control

Actions use Velora’s existing approval policy. Taking over blocks agent actions until you return control. Screenshots become local task artifacts; their image bytes are excluded from activity logs. This feature does not grant unrestricted control over your own desktop.

Provider availability, model support and quota still apply. Velora Auto’s hosted Brain path does not yet support desktop screenshot transport; select a connected vision-capable provider for visual work. No paid model fallback is enabled automatically.