Why I gave Claude Code a computer instead of building another IDE

For years, AI coding tools have simply placed a chat window adjacent to your editor, treating it as sufficient. However, the author desired a more robust solution that operates like Claude Code does on a server - providing a real computer environment for the AI agent to work within. Superagent is a Mac application that accomplishes this by offering Claude Code (or any designated agent) access to a genuine environment, rather than just a text box.

This includes a functional browser capable of navigating, clicking, typing, and reading the Document Object Model (DOM); an iOS Simulator window for installing apps, tapping through interfaces, and screenshotting to verify UI changes; and a relay pairing the Mac with a smartphone, enabling the agent to continue working while the user remains nearby.

The primary objective behind building Superagent was not to create a more sophisticated chat window, but to establish an optimal platform for the AI agent to execute tasks. The majority of difficulties encountered during development stemmed from the environment itself, such as synthetic file inputs failing to persist through web form upload components, elements appearing in the DOM but absent from the accessibility tree, and simulator states deviating from the actual screenshot.

Addressing these issues is what ultimately renders the agent reliable to tackle assigned tasks. Superagent is constructed as a trio of components: an Electron desktop application, a SwiftUI iOS companion, and a compact Cloudflare Worker relay facilitating communication between the two while imposing per-address rate limits to prevent misuse via a lost phone.

The source code is available at https://peerlist.io/pungme/project/superagent-for-mac. The author welcomes inquiries regarding the browser automation strategy, simulator manipulation, or the relay's pairing and security protocols in the comments section.