Architecture

Clientless has one architectural boundary: capabilities describe what can be done, and hosts assemble those capabilities into personal interfaces.

developer capabilities
        ↓
action, data, permission, and view contracts
        ↓
json-render spec or stream
        ↓
MCP Apps resource + host bridge
        ↓
registry-backed renderer
        ↓
personal interface in chat, web, mobile, desktop, widget, or static output

Capabilities

A capability is a useful part of a product exposed as something a host can call and render.

It can include:

• actions,
• data queries,
• generated views,
• decision support,
• approval flows,
• validation rules,
• long-running tasks,
• saved interface state.

Capabilities can be implemented with any service architecture. That implementation detail stays behind the contract.

Interface contract

The interface contract is json-render.

A spec describes components, props, state bindings, visibility, and actions. A stream progressively patches that spec so a personal interface can appear while data, tools, and model output resolve.

MCP Apps boundary

MCP Apps provides the standardized application resource model. A capability can return interactive UI, and a host can render it in a controlled environment while mediating calls safely.

Hosts

Hosts assemble and render capabilities. They provide:

• transport,
• resource loading,
• sandboxing and security policy,
• platform context,
• action forwarding,
• a component registry for the target surface,
• save and recall behavior for personal interfaces.

The more a host can assemble safely, the less users need fixed applications as destinations.