The Problem: Feature Request Bottlenecks

Traditional software development follows a rigid, top-down model: developers build a "blessed" version of an app, and users submit feature requests that often languish in Jira. When developers attempt to solve this via plugin architectures, they often trigger years-long rewrites that stall product progress. Varda argues that personal AI codegen breaks this model by allowing users to have their own agents modify the app specifically for their needs, keeping the core codebase clean while satisfying individual requirements.

The Gadgets Architecture

To enable this safely, Varda proposes "Gadgets"—a platform built on Cloudflare Workers that treats applications like documents in an office suite. Each gadget is a single, isolated instance of an app.

• Platform-Level Sharing: Because the platform manages sharing and access control, individual apps cannot misconfigure permissions.
• Safe Extensibility: Users can prompt an LLM to add features (like custom SVG rendering or specific UI formatting) directly to their instance.
• Security through Isolation: The UI runs in a null-origin iframe, and server code runs in a dynamic worker sandbox. Communication is restricted to a Cap'n Proto RPC session between the client and server. This architecture renders XSS bugs effectively harmless, as the code has no access to external data or cookies to leak.

Infrastructure and Execution

Unlike traditional cloud setups that rely on containers and databases, Gadgets uses Cloudflare's workerd runtime and Durable Objects. This approach allows for:

• Local Execution: The entire environment can run locally on a developer's machine, independent of network availability.
• Self-Hosting Potential: Because the runtime is open-source, users can eventually host their own gadgets for tasks like home automation or local data processing, moving away from the "ivory tower" model of centralized cloud distribution.